Multi-source data driven ecological protection and restoration method and system

Abstract

The invention discloses a multi-source data-driven ecological protection and restoration method and system, and aims to solve the technical problems of contradiction between wetland and forest environment monitoring precision and cost, poor data transmission adaptability, and weak analysis model over-fitting and generalization ability. According to the method, a heterogeneous wireless ad hoc network driven by multi-source data is constructed, and a wireless sensor network and a Mesh network are fused, so that efficient sensing and reliable transmission of wetland and forest land environments are realized; low-cost sensor data is processed by adopting a machine learning and optimization theory, a lightweight environment intelligent detection model is established, an environment-sensitive wireless sensor MAC protocol is optimally designed, and the data transmission efficiency and reliability in a complex environment are improved. According to the method, the monitoring cost can be remarkably reduced, the accuracy and efficiency of ecological environment assessment and pest control and prevention are improved, technical support is provided for ecological protection and restoration, healthy tourism development and digital forest protection, and the method has important application value and social benefits.

Description

Technical Field

[0001] This invention relates to the field of ecological protection and restoration technology, specifically to a multi-source data-driven method and system for ecological protection and restoration. Background Technology

[0002] Digitalization plays a crucial role in ecological protection and restoration. Ecological prosperity leads to civilizational prosperity; nature is the fundamental condition for human survival and development, and ecological civilization construction is a fundamental plan related to social development. Territorial ecological protection and restoration is an important component of ecological civilization construction. Recognizing and promoting the high-quality development of the digital economy from a holistic perspective, fostering the deep integration of digital technology and the real economy, advancing digital industrialization and industrial digitalization, and comprehensively empowering economic and social development have become a general consensus. Therefore, building a data-driven model to assist staff in carrying out ecological protection and restoration projects can effectively reduce labor costs and has significant application value and social benefits for local environmental protection and economic development.

[0003] In recent years, great importance has been attached to the ecological protection and restoration of wetlands and forests. Measures such as establishing ecological protection red lines, implementing ecological restoration projects, and constructing ecological monitoring systems have been implemented to ensure the ecological security of parks. Pine wilt disease, due to its high pathogenicity, rapid spread, and difficulty in eradication, causes devastating damage to pine trees and is known as the "cancer" of pine trees. It has been listed as an important quarantine target both internationally and domestically. Pine wilt disease is mainly transmitted naturally through the pine sawyer beetle. The pine sawyer beetle, through its activities such as flying or crawling, spreads the pine wilt nematode to new pine trees, allowing the nematode to migrate widely throughout the pine forest ecosystem.

[0004] Currently, the main methods for controlling pine wilt disease include the following four categories: 1. Using traps to kill pine sawyer beetles; 2. Spraying pine sawyer beetles with thiamethoxam; 3. Using pine wilt disease injection immunization agents; 4. Organizing the unified removal and treatment of pine wilt disease-infected dead trees owned by village collectives or farmers.

[0005] Environmental monitoring and data analysis are crucial for ecological protection and restoration. However, achieving efficient and reliable multi-source data-driven monitoring systems in complex wetland and forest environments presents numerous technical challenges. Extensive research and analysis revealed that multi-sensor wireless ad hoc network transmission in wetland and forest environments requires a key solution: integrating protocol stack layers with ecological and environmental information to achieve environmental sensitivity of the protocol. Relative and representative (R / S) analysis, a crucial tool in time series research, has been used to analyze the temporal characteristics of vegetation and tree growth. In line with the research objectives of this project, the following sections will describe the current research status in environmental monitoring systems, wireless sensor network MAC protocols, reliable routing, transmission control, and time series analysis.

[0006] In the field of environmental monitoring research, many researchers have proposed different solutions. In fact, in complex ecological environments, wireless sensor network monitoring systems are more affected by environmental factors and face many problems that urgently need to be solved, such as the adaptability of MAC and routing protocols.

[0007] As can be seen from the above, existing routing protocol research has explored connectivity and reliability in depth, but it rarely addresses how to integrate environmental information, node location, and MAC information to achieve multi-source data-driven operation. To ensure correct and reliable data transmission, transport layer protocols are also crucial. Existing research has yielded numerous transmission control mechanisms for wireless sensor networks and ad hoc networks, but most designs are independent of routing protocols, resulting in performance inefficiencies. Some research suggests that transmission control should cooperate with routing.

[0008] In data analysis research, current methods all involve comprehensive sampling at the sample site, establishing datasets by collecting modeling and validation samples, conducting correlation analysis between vegetation and tree parameters and environmental and climatic factors, and ultimately constructing regression prediction models. Regression analysis typically requires large sample sizes; if the number of data samples is insufficient, the constructed regression model is prone to inaccurate regression coefficient estimation and overfitting, ultimately leading to significant deviations in prediction results. To overcome the shortcomings of insufficient sample data, this invention employs time series methods for prediction. By analyzing the characteristics of plant growth patterns, trends, and periodic changes in wetlands and forests, it reduces the possibility of excessive fitting errors due to small sample sizes. Regression analysis suffers from the problem of small training sample data leading to model overfitting; nonlinear time series models effectively compensate for the shortcomings of regression models, but the fitted mathematical model is a power function, which is difficult to apply to the complex and variable ecological environments of wetlands and mountain forests, resulting in weak model generalization ability. Therefore, this invention adopts a time series prediction method, simultaneously solving the problems of overfitting and weak generalization ability of traditional models.

[0009] In summary, establishing multi-source data-driven heterogeneous wireless ad hoc networks in wetland and forest environments to achieve efficient and reliable monitoring data transmission schemes and data analysis and prediction models faces many technical challenges, making their research and optimization essential. Summary of the Invention

[0010] In view of the prior art, the object of the present invention is to address the shortcomings of the prior art in the following two aspects:

[0011] (1) How to use inexpensive sensors to achieve environmental monitoring with acceptable accuracy, and to use UAV remote sensing data to perform time-series analysis and prediction of wetland and forest environmental conditions and to detect pine wilt disease plants;

[0012] (2) How to design an environment-sensitive wireless sensor network MAC protocol, and how to integrate environment, cross-layer and geographical location to design an efficient and reliable multi-source data-driven mode.

[0013] To achieve the above objectives, the technical solution adopted by this invention is as follows: a multi-source data-driven ecological protection and restoration method, comprising the following steps: Step 1: Constructing a multi-source data-driven wireless sensor network monitoring system, integrating wireless sensor networks and Mesh networks to form a heterogeneous wireless self-organizing network, deploying multiple types of sensors to collect fine-grained ecological environment information of wetlands and forests, and simultaneously acquiring remote sensing image data through a UAV equipped with a binocular depth camera; Step 2: Based on the data collected by the wireless sensor network monitoring system, designing a calculation model for the health and wellness index of environmentally sensitive wetlands and forests, standardizing the data, calculating climate comfort indicators and performing principal component analysis, and finally realizing the calculation and classification of the health and wellness index; Step 3: Integrating the environmental data and remote sensing image data to form a time series, constructing a pine wilt disease prediction model based on R / S analysis and Cauchy kernel correlation vector machine, and combining UAV AI recognition and geographic information system to achieve accurate location of diseased pine trees and epidemic prediction, thereby realizing the prediction of pine wilt disease and accurate location of diseased plants; Step 4: Optimize the MAC protocol, routing protocol, and transmission control protocol of the wireless sensor network, integrate environmental perception data and node location information to achieve cross-layer collaboration, and build a heterogeneous wireless ad hoc network that integrates the wireless sensor network and the Mesh network to ensure the efficiency and reliability of data transmission; Step 5: Deploy the heterogeneous wireless ad hoc network in wetlands and woodlands for long-term testing, and optimize the model and protocol based on the test results.

[0014] As a further feature of this scheme, the multiple types of sensors mentioned in step one include at least two of the following: wind speed sensor, air pressure sensor, light sensor, sound wave sensor, humidity sensor, temperature sensor, raindrop sensor, negative ion sensor, PM2.5 sensor, noise sensor, carbon dioxide sensor, and ultraviolet radiation sensor; the drone is a Phase One X450 drone, which is equipped with an Intel d435i binocular depth camera and an NVIDIA Jetson TX2 embedded processing platform; the embedded processing platform is used to pre-train a YOLOv5 target detection model to identify pine tree diseases and pests in remote sensing images. The YOLOv5 model improves detection accuracy by adjusting the anchor structure, optimizing the step size, adding detection layers, and using a weighted bounding box fusion algorithm.

[0015] As a further feature of this scheme, the construction process of the health and wellness index calculation model in step two includes: using the formula Calculate the climate comfort index S, where T is air temperature, RH is relative humidity, and V is wind speed. Standardize the indicators for negative ions, PM2.5, noise, carbon dioxide, ultraviolet radiation index, and climate comfort index. For indicators where a higher value is better, use the formula... Standardization is performed; for indicators where smaller is better, a formula is used. Standardization is carried out, among which Perform principal component calculations.

[0016] F1 = Comfort * 0.826 - UV * 0.78 + Negative Ions * 0.739 - Noise * 0.721

[0017] F2 = PM2.5 * 0.924

[0018] F3 = carbon dioxide * 0.948

[0019] All parameters are standardized values;

[0020] The Forest Health and Wellness Index (FHCI) is calculated as follows: F1 × 0.519255438908815 + F2 × 0.260347671283085 + F3 × 0.2203968898081. Based on the FHCI value, the health and wellness index is divided into four levels: FHCI ≥ 0.60 is Level 1, 0.20 ≤ FHCI < 0.60 is Level 2, 0.0 ≤ FHCI < 0.20 is Level 3, and FHCI < 0 is Level 4.

[0021] As a further feature of this scheme, the construction process of the pine wilt disease prediction model in step three includes:

[0022] (1) The Hurst exponent H of pine diameter growth data was calculated using the R / S analysis method to analyze the temporal characteristics of pine diameter growth;

[0023] (2) A regression function is constructed based on the Cauchy kernel correlation vector machine, and a pine wilt disease prediction model is established using the Hearst exponent H time series data as training samples. The Cauchy kernel function is... , where σ is the Cauchy kernel parameter.

[0024] 1. As a further feature of this scheme, the process of calculating the Hearst exponent H using the R / S analysis method includes: segmenting the time series to obtain multiple sub-intervals. ; Calculate each subinterval Cumulative deviation :

[0025]

[0026] in, Reflects the cumulative shift of data points relative to the mean. This is time series data. sub-interval The mean; calculate the mean of each subinterval. range and standard deviation

[0027]

[0028] Calculate the rescaled range ratio

[0029]

[0030] For different window lengths ,Establish and linear relationship Where H is the Hurst exponent and C is the intercept term; the training data in the time-series experimental data for measuring the diameter growth of pine trees. The training sample set is constructed as follows:

[0031] ,

[0032] In the formula: X is the input sample set, and Y is the corresponding output set; a Cauchy kernel function is constructed, and a regression function for a Cauchy kernel correlation vector machine for predicting the time series of pine tree diameter growth is established based on the training sample set:

[0033]

[0034] In the formula: Indicates weight, This represents a mapping function.

[0035] As a further feature of this scheme, the MAC protocol of the wireless sensor network in step one is an environment-sensitive protocol. It integrates the results of lightweight environmental perception and prediction models, optimizes and adjusts the MAC layer's contention and avoidance mechanisms for the channel, and dynamically adjusts the MAC protocol's sleep and synchronization strategies to adapt to the environmental characteristics of wetlands and woodlands with high humidity and multiple interferences. The routing protocol is a reliable routing protocol that integrates cross-layer and geographical location information. It uses the cross-layer interface extended by the protocol stack to obtain MAC layer information and achieve efficient and reliable data transmission.

[0036] This invention also provides a multi-source data-driven ecological protection and restoration system, comprising: an S1 data acquisition module, including a multi-source sensor network and a UAV remote sensing unit, wherein the multi-source sensor network is deployed in wetland and forest environments to collect fine-grained ecological environment information, and the UAV remote sensing unit is used to acquire remote sensing image data of wetlands and forests; an S2 data transmission module, which is a heterogeneous wireless ad hoc network integrating wireless sensor networks and mesh networks, used to achieve reliable transmission of data collected by the data acquisition module; an S3 data processing and analysis module, deployed on an embedded platform, including a health and wellness index calculation unit and a pine wilt disease prediction unit; and an S4 control and optimization module, used to optimize and adjust the communication protocol of the data transmission module and the model of the data processing and analysis module.

[0037] Prior to this, the UAV remote sensing unit also includes a navigation module, which adopts a dual-stream deep network architecture, combines real-time sensing information and historical navigation data, and outputs navigation decisions through SoftMax normalization to achieve autonomous environment exploration and target positioning.

[0038] Prior to this, the data processing and analysis module uses an incremental learning method to dynamically adjust the parameters of the environmental perception model to adapt to the dynamic changes in wetland and forest environments. The system supports pre-evaluation of protocol performance on the NS3 simulation platform and continuously optimizes the model and protocol through outdoor measured data.

[0039] Priority is also given to the ground control module, which communicates bidirectionally with the data transmission module to receive data processing results and send control commands to the UAV remote sensing unit, thereby enabling the UAV to navigate and explore autonomously.

[0040] The system and method of this invention are of great scientific significance, application value and innovation for building a multi-source data-driven wireless sensor network monitoring system in ecological environments such as wetlands and forests. The key technology for realizing the monitoring system lies in how to effectively deploy a large-scale wireless sensor network to acquire time-series data in wetlands and forests to form key decision information.

[0041] Furthermore, the communication distance of wireless sensor network nodes is generally short, and the monitoring range covered by a single node or cluster is limited. Data transmission between multiple clusters or different wireless sensor networks requires a network with a wider communication range. Therefore, this invention considers integrating wireless sensor networks and mesh networks to construct a heterogeneous wireless self-organizing network, forming a wetland and forest monitoring network covering the entire region. The wetland and forest monitoring network is optimized using a multi-source data-driven model, and long-term collection of ecological and environmental data is conducted to form a time-series analysis model. The temporal variation characteristics of wetlands and forests are core parameters for evaluating environmental quality and the effectiveness of management techniques. Systematically analyzing the growth patterns of green plants and pine forests and adopting scientific management strategies can significantly optimize plant growth performance, thereby achieving ideal economic output.

[0042] This invention addresses the needs arising from wetland monitoring and pine pest and disease control. Facing challenges such as high maintenance costs, harsh outdoor climates, scattered forest areas, and inconvenience of manual inspections, it starts from the digital technology requirements of ecological environment protection and restoration work. It integrates wireless sensor networks and mesh networks to construct a multi-source data-driven heterogeneous wireless ad hoc network, solving technical problems related to the perception and transmission of ecological environment information in wetlands and forests. This invention closely focuses on key technologies such as wetland and forest environment monitoring and pine wilt disease prediction, and can generate significant application value and social benefits for socio-economic development and ecological construction. Attached Figure Description

[0043] Figure 1 This is a block diagram of the technical roadmap for the multi-source data-driven ecological protection and restoration method and system of the present invention.

[0044] Figure 2 This is a schematic diagram of the Phase One X450 UAV used in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the hardware communication architecture used in the embodiments of the present invention.

[0046] Figure 4 This is an example of an ecological environment data sample from UAV remote sensing imagery of the present invention.

[0047] Figure 5 This is a schematic diagram of the time-series prediction process for pine tree diameter growth based on CRVM according to the present invention.

[0048] Figure 6 This is a flowchart of the deep learning algorithm for identifying pine wilt disease in this invention.

[0049] Figure 7 This is a schematic diagram of the architecture of the YOLOv5 algorithm used in this invention.

[0050] Figure 8 This is a schematic diagram of the modified anchor structure of the present invention.

[0051] Figure 9 This is a schematic diagram of the step size experimental design of the present invention.

[0052] Figure 10 This is a schematic diagram of the detection layer of the present invention.

[0053] Figure 11 This is a schematic diagram comparing NMS and WBF of the present invention.

[0054] Figure 12 This is a schematic diagram illustrating the autonomous environment exploration of the present invention. Detailed Implementation

[0055] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] In addition, it should be noted in advance that:

[0057] Efficient sensing, reliable transmission, and analysis / prediction of wetland and forest environmental data are crucial for green economic development and ecological environmental protection. However, current sensing and transmission technologies used in outdoor environments face numerous new challenges in wetland and forest environments, such as the impact of harsh environments like high humidity, heavy rainfall, and typhoons on sensing and transmission modules (information perception accuracy, transmission correctness, and power consumption). This is the starting point for this invention. Furthermore, the deployed multi-sensor heterogeneous wireless ad hoc network can lead to model overfitting during data analysis. While nonlinear time-series models effectively compensate for the shortcomings of regression models, they are difficult to apply to complex and variable ecological environments, resulting in weak model generalization ability. In this invention, to achieve efficient sensing, reliable transmission, and analysis / prediction of wetland and forest environmental data, a multi-source data-driven wireless sensor network monitoring system is established. An environmentally sensitive wetland and forest health and wellness index calculation model is designed, and a time-series-based pine wilt disease prediction model is constructed.

[0058] The specific content includes the following three aspects:

[0059] Research on Multi-Source Data-Driven Wireless Sensor Network Monitoring System

[0060] In existing research, accurate environmental detection typically relies on high-precision sensors, which comes at a high cost. This high cost leads to sparse sensor deployment and coarse-grained environmental data that fails to comprehensively reflect environmental characteristics. To achieve acceptable monitoring accuracy while reducing deployment costs, this invention considers using machine learning methods and optimization theory to process and analyze data collected by large-scale, low-cost sensors. A lightweight intelligent detection model for wetland and forest environments is defined and established, validated using machine learning and optimization algorithm libraries, and deployed and tested in a wireless sensor network in wetlands and pine forests. In wireless sensor networks, the quality of the MAC layer protocol directly determines the efficiency and reliability of data transmission and also affects the overall network lifetime. Existing research on wireless sensor MAC layer protocols mainly focuses on reducing energy consumption, rarely considering the characteristics of the physical environment in which they operate. This leads to issues with sensing efficiency and transmission reliability in wetland and forest environments. To address this problem, this invention considers the environmental characteristics of sensor detection, integrates the intelligent environmental detection model proposed above, and optimizes the transmission control, sleep measurement, and synchronization mechanisms of the MAC layer protocol to design and implement a reliable MAC protocol for environmentally sensitive wireless sensors.

[0061] Research on Calculation Model of Health and Wellness Index for Environmentally Sensitive Wetlands and Forest Lands

[0062] The ecological environment health and wellness index can help assess the suitability of wetlands and scenic areas for health and wellness tourism, thereby enabling the formulation of corresponding policies and measures to promote its development. Furthermore, the index can help tourists choose suitable health and wellness tourism destinations, enhancing their travel experience and health benefits. In addition, the index can promote regional sustainable development. By improving the ecological environment and enhancing the suitability for health and wellness tourism, more tourists can be attracted, thus driving regional economic development. This study optimizes and designs a health and wellness index calculation model suitable for wetland and forest environments. A multi-source data-driven wireless sensor network monitoring system is used to collect real-time wetland and forest environmental data, including negative ions, PM2.5, noise, carbon dioxide, temperature, humidity, wind speed, and ultraviolet radiation index. The data is standardized, and a climate comfort index is calculated. Principal component analysis is performed using the climate comfort index, ultimately achieving the calculation and effective classification of the wetland and forest health and wellness index.

[0063] Research on Time Series-Based Prediction Model for Pine Wilt Disease

[0064] Controlling pine wilt disease in pine forests is challenging. Relying solely on frontline rangers to patrol and manage large areas of dead pine trees, effectively curbing the spread of the disease, only represents the first step in prevention and control. "Digital forest protection," or intelligent management, is key to achieving precise and efficient prevention and control. By deploying a multi-source data-driven wireless sensor network monitoring system, the collected sensor data can form a multi-source data-driven model, effectively constructing discrete data time series for predicting pine wilt disease. Addressing the safety hazards of manual inspection, technologies such as UAV remote sensing data acquisition, AI recognition, and geographic information systems are used to assist in monitoring pine pests and diseases. UAV aerial remote sensing monitoring technology has been developed to achieve precise location of diseased pine trees. Simultaneously, time series data is used to identify pine wilt disease infestation factors and predictive models, completing data management informatization and effectively solving the problems of high difficulty and time-consuming manual monitoring, truly achieving comprehensive coverage of investigation and monitoring work.

[0065] This invention addresses ecological protection and restoration issues, defining and establishing a lightweight environmental perception and prediction model driven by multi-source data. It also designs and implements an environmentally sensitive health and wellness index calculation model that meets the environmental transmission needs of wetlands and forests, as well as a time-series-based prediction model for pine wilt disease. The specific technical approach of the project is as follows: Figure 1 As shown.

[0066] like Figure 1 The technical route shown in this invention is divided into 5 stages:

[0067] I. We use a multi-source data-driven wireless sensor network to sense wetland and forest environmental information and perform time series analysis. We analyze the performance of the wireless sensor network MAC protocol, wireless ad hoc network routing and transmission protocol in a network simulation environment.

[0068] II. Propose and design a health and wellness index calculation model applicable to wetland and forest environments that are sensitive to environmental conditions, calculate the ecological environment health and wellness index in real time and classify it effectively;

[0069] 3. Using remote sensing image data and multi-sensor time series data, identify the factors that cause pine wilt disease and construct a pine wilt disease prediction model.

[0070] Fourth, the above model will be further implemented and optimized on an embedded system, and analyzed and debugged in an outdoor experimental network environment;

[0071] 5. Deploy wireless sensor network nodes in actual wetlands and woodlands to build a heterogeneous wireless ad hoc network, and conduct long-term testing and optimization. Optimize the model and protocol based on the results.

[0072] The technical details of the above five stages are as follows:

[0073] (1) Measurement and Evaluation: For wetland and forest environmental information perception, multi-sensor detection information is integrated, such as wind speed, air pressure, light intensity, sound waves, humidity, temperature, and raindrops, to collect fine-grained ecological and environmental information. To measure and analyze the environmental monitoring capabilities of conventional wireless sensor networks, this invention deploys experimental networks in wetlands and forests for environmental monitoring and analyzes the performance of MAC layer and network layer protocols (such as throughput, latency, packet loss rate, and node power consumption). In addition, for the micro-performance analysis of wireless sensor network protocols, the project uses the network simulation platform (NS3) to set up experimental scenarios with wetland and forest environmental characteristics, and evaluates the performance of different protocols under varying conditions such as node size and traffic load.

[0074] This invention uses a Phase One X450 drone equipped with an Intel d435i binocular depth camera as the basic experimental platform, such as... Figure 2 As shown, the D435i camera has four circular holes on the front. From left to right, the first and third are infrared sensors (IRStereo Camera); the second is an infrared laser emitter (IR Projector); and the fourth is a color camera (color sensor). The camera can capture images up to 10 meters away, and the video transmission rate can reach 90fps.

[0075] The onboard image processing board used is the Nvidia TX2 embedded platform, designed for unmanned intelligent applications. It is a modular AI supercomputer with an NVIDIA Pascal™ GPU featuring 256 CUDA cores. The CPU has six cores: a dual-core Denver2 processor and a quad-core ARM Cortex-A57. Its powerful performance and compact size make it ideal for intelligent edge devices such as robots, drones, and smart cameras. After training the YOLOv5 model on a virtual machine, the model is imported to the TX2 processing board via an SD card. The trained model is then run through the corresponding path to generate the processing results.

[0076] Figure 3 The communication architecture of the hardware platform used in this invention is described in detail. In the figure, the binocular camera, vision processing board, flight controller, and airborne data transmission are all fixedly installed on the UAV platform and rise into the air together with the UAV platform during operation; the ground data transmission is installed on the ground station and realizes bidirectional data communication with the airborne data transmission, and is placed on the ground together with the ground station during operation.

[0077] The entire system's working process and data flow are described below:

[0078] 1. The binocular camera acquires image data of the area to be detected and recognized, preprocesses it into depth image data, and outputs it to the vision processing board;

[0079] 2. The target detection algorithm and UAV exploration agent algorithm running on the airborne vision processing board can process the depth image data sequentially and finally output a target position information to the flight control.

[0080] 3. After receiving the target location information, the flight control system will control the UAV platform to fly over the target location;

[0081] 4. To achieve the detection of a point, after the UAV platform flies over the target, the system repeats (1) to (3) until all points to be detected have achieved the above operation.

[0082] (2) Protocol Modeling: Based on the measurement and evaluation in (1), a lightweight perception and prediction model for complex wetland and woodland environments is defined and established. An environmentally sensitive wireless sensor network MAC layer protocol and a routing and transmission control protocol that integrates cross-layer and geographic location are proposed, and modeled, simulated and implemented. Among them, machine learning and optimization methods are used for modeling the environmental perception and prediction model.

[0083] Assumption These represent the pasture environmental information detected at specific times by sensors such as air pressure, light intensity, sound, humidity, wind speed, temperature, and raindrops. Given all environmental information detected by the sensors, where n is the number of sensor types, a learning algorithm can be established using machine learning combined with optimization theory. Online learning yields perception and prediction models. .

[0084] In practical implementation, since sensor detection data is periodic and continuous, an incremental learning method is used to continuously adjust the learning model. In the design of MAC protocols for environment-sensitive wireless sensors, the results of lightweight environmental perception and prediction models are integrated to optimize and adjust the MAC layer's channel contention and avoidance mechanisms, and to dynamically adjust the MAC protocol's sleep and synchronization strategies. In routing protocols that integrate cross-layer and geographic location information, MAC layer information is fully considered.

[0085] In terms of specific design and implementation, based on geographic location routing, the cross-layer interface extended by the protocol stack is used to obtain MAC layer information (obtained through the learning model defined in this invention) to achieve reliable routing across layers and geographic locations, so as to meet the requirements of wetland and forest environment monitoring for transmission efficiency and reliability.

[0086] Since this invention is a heterogeneous network consisting of two parts: a wireless sensor network (such as an extended TinyOS) and a mesh network (such as an extended OpenWRT), data interaction between the two different networks needs to be implemented to ensure normal data transmission. To meet the requirements of low power consumption and data transmission distance in marine environments, the MAC layer of the mesh network nodes (gateways) utilizes 802.11ah.

[0087] This invention selects photos from UAV remote sensing images of wetlands and pine forests as a sample library, with 70% of the images used as the training set and the remaining 30% as the test set. Professional labeling software is used to annotate the remote sensing information of wetlands and forests in the UAV images. The coordinates of the upper left and lower right corners of the bounding boxes are recorded in an Extensible Markup Language (EXPLAIN) file, such as... Figure 4 As shown.

[0088] Due to the limited number of ecological environment data samples in UAV remote sensing imagery, this invention supplements the training with data downloaded from publicly available datasets. Simultaneously, this invention performs random cropping, rotation, scaling, and horizontal flipping operations on the data to generate multiple similar images. Data augmentation not only compensates for the insufficient data volume but also effectively reduces overfitting, making the model more suitable for new samples and enhancing its generalization ability. The UAV images are ultimately compiled into a dataset in a visual object class format for pre-training the deep learning model.

[0089] (3) Model design: Based on (1) measurement and evaluation and (2) protocol modeling, various ecological environment data and UAV remote sensing data are obtained, and an environmentally sensitive health index calculation model and a time series-based pine wilt disease prediction model are designed and implemented.

[0090] ① Calculation model of environmentally sensitive health and wellness index

[0091] By analyzing data on negative ions, PM2.5, noise, carbon dioxide, temperature, humidity, wind speed, and ultraviolet radiation index, a climate comfort index can be calculated using data on temperature, humidity, and wind speed. The climate comfort formula is as follows:

[0092] (1)

[0093] In the formula: S is the overall comfort index, T is the air temperature, RH is the relative humidity, and V is the wind speed.

[0094] Indicators are categorized into two types based on their nature: the larger the better and the smaller the better. For indicators of the larger the better type, standardization is performed using formula (2); for indicators of the smaller the better type, standardization is performed using formula (3).

[0095] (2)

[0096] (3)

[0097] in, .

[0098] If the monitored index value exceeds the max and min values ​​in the table below, replace the values ​​in the table below with the measured max and min values, and then perform standardization.

[0099] Table 1. Comparison of Standardized Max and Min Values ​​of Indicators

[0100] standardization negative ions PM2.5 noise carbon dioxide UV Index S Climate Moderation max 39901.07463 82.19265419 69.40291262 520.4753521 2.486206897 15.12326389 min 28119.5527 24.73240418 41.72080444 399.2670509 0 1.04625 max-min 11781.52192 57.46025001 27.68210818 121.2083012 2.486206897 14.07701389

[0101] Principal component calculations are then performed:

[0102] F1 = Comfort * 0.826 - UV * 0.78 + Negative Ions * 0.739 - Noise * 0.721

[0103] F2 = PM2.5 * 0.924

[0104] F3 = Carbon dioxide * 0.948

[0105] All the above parameters use standardized values.

[0106] The forest health and wellness index can be calculated using principal components F1, F2, and F3:

[0107] FHCI=F1*0.519255438908815+F2*0.260347671283085+F3*0.2203968898081

[0108] Based on the FHCI value, the forest health index can be divided into 4 levels. FHCI ≥ 0.60 is level 1 (very good, very beneficial); 0.20 ≤ FHCI < 0.60 is level 2 (good, beneficial); 0.0 ≤ FHCI < 0.20 is level 3 (average, normal); FHCI < 0 is level 4 (poor, not beneficial).

[0109] ② Time-series-based prediction model for pine wilt disease

[0110] RVM is a Bayesian framework implementation of SVM, providing a probabilistic interpretation of the SVM output. The SVM prediction model is estimated based on the following form:

[0111] (4)

[0112] In the formula: Indicates weight, Indicates deviation, For kernel function, This is the input vector.

[0113] The RVM modeling process begins with a set of input variables and their corresponding target output values. Given a set of data... ,in, It is a set of input vectors. It is the corresponding output target.

[0114] Given the regression function of RVM :

[0115] (5)

[0116] In the formula: Indicates weight, This represents a mapping function.

[0117] The likelihood function of the dataset is described as follows:

[0118] (6)

[0119] (7)

[0120] In the formula: Represents the kernel function. This indicates the output value containing Gaussian noise. This represents the noise variance.

[0121] Introduce higher-order hyperparameters to constrain the prior condition distribution of the weight vector:

[0122] (8)

[0123] In the formula: Represents a hyperparameter vector. This represents the Gaussian conditional probability distribution.

[0124] The non-informative prior distribution is as follows:

[0125] (9)

[0126] (10)

[0127] The marginal likelihood function of the hyperparameters is obtained by the following formula:

[0128]

[0129] (11)

[0130] In the formula: , It is a diagonal matrix.

[0131] The kernel function uses the Cauchy kernel. :

[0132] (12)

[0133] In the formula: This represents the Cauchy kernel parameters.

[0134] Assuming the pine trees are currently infected with pine wilt disease, the Hurst index (H-value) can be used to measure the diameter growth parameter. The H-value based on pine diameter growth data can be calculated using the R / S analysis method. As a key tool in time series research, R / S analysis was used in this study to analyze the temporal characteristics of pine diameter growth, and two growth prediction models, linear and power-law, were constructed based on the analysis results. The calculation process is shown below:

[0135] For each sub-interval Calculate cumulative deviation :

[0136] (13)

[0137] In the formula: Reflects the cumulative shift of data points relative to the mean. It is a time series. This represents a sub-interval of the time series.

[0138] Range of each sub-interval for It reflects the range, which characterizes the fluctuation range of the sequence within that window. For each sub-interval... Calculate the standard deviation :

[0139] (14)

[0140] Standard deviation is used to standardize the range, eliminating the effects of data scaling. For each sub-interval... Calculate the rescaled range ratio This ratio quantifies the degree to which the fluctuation of a sequence deviates from random fluctuations.

[0141] For different window lengths , and Linear relationship:

[0142] (15)

[0143] in, The Hearst exponent. This is the intercept term.

[0144] Given training data from time-series experimental data measuring the H value of pine tree diameter growth. The training sample set is constructed as follows:

[0145] , (16)

[0146] In the formula: X is the input sample set, and Y is the corresponding output set.

[0147] Construct the Cauchy kernel function and, based on the training sample set, establish a Cauchy kernel relevance vector machine (CRVM) regression function for time-series prediction of pine tree diameter growth:

[0148] (17)

[0149] In the formula: Indicates weight, This represents a mapping function.

[0150] We used CRVM, a time series prediction tool for pine tree diameter growth, to perform time series prediction of the H value, which measures pine tree diameter growth, obtain the predicted H value, and calculate its relative error.

[0151] The CRVM-based time-series prediction process for pine tree diameter growth is as follows: Figure 5 As shown.

[0152] (4) Analysis and Measurement: Based on the established model and designed protocol, implement the corresponding modules on the network simulation platform, and establish experimental scenarios with wetland and forest environment characteristics. Perform performance analysis on the model and protocol, and determine whether to optimize key model parameters based on the analysis results. For the established and optimized model and protocol, design and implement it in the protocol stack of the wireless sensor network operating system (such as TinyOS), and perform measurement and analysis in an experimental heterogeneous wireless ad hoc network (integrating wireless sensor network and Mesh network) environment. For different network sizes, environments and network loads, determine whether to optimize the key parameters of the model based on the results.

[0153] Target detection of pine wilt diseased plants is highly challenging, with problems such as low training data volume, highly imbalanced datasets, and frequent changes in target position and scene due to shaking during aerial photography by airborne optoelectronic pods.

[0154] This invention uses YOLOv5 as the baseline algorithm and proposes various optimization strategies to address problems encountered in object detection. The overall algorithm flow is as follows: Figure 6 As shown.

[0155] A deep learning network, YOLOv5, for object detection is used to detect objects on the labeled training set images. The overall architecture of YOLOv5 is as follows: Figure 7 As shown.

[0156] YOLOv5 is the network with the smallest depth and feature map width in the object detection series, achieving accuracy comparable to YOLOv4, but with a model size nearly 90% smaller. YOLOv5 converges quickly on multiple datasets, offers high model customizability, and is an excellent lightweight network. The relevant source code can be found at https: / / github.com / ultralytics / YOLOv5. This invention implements YOLOv5 based on the PyTorch framework. The added Focus operation transforms image slices into feature maps, significantly improving data processing speed. Two CSP structures are used in the backbone extraction network. Taking the YOLOv5 network as an example, the CSP1_X structure is applied to the backbone network, while the other CSP2_X structure is applied to the Neck. For the neck part, an FPN+PAN structure is chosen. The CSP2 structure, inspired by CSPnet, is adopted to enhance the network's feature fusion capabilities. During the prediction process, GIOU_Loss is used to address the issue of IoU Loss where the loss is 0 when B and G do not intersect. This allows the predicted bounding box to more accurately detect forest pests and diseases.

[0157] To further improve detection accuracy, this invention improves YOLOv5 in four aspects: step size, anchor structure, detection layer, and fusion frame.

[0158] 1. Modify the anchor structure

[0159] Unlike the large-scale, multi-class dataset of COCO, and given that forest pest and disease data are mostly below 0.3m, the anchor structure was adjusted to make the model learn more easily and converge better and faster. For example... Figure 8 As shown.

[0160] 2. Modify step size

[0161] Stride represents the step size of the filter in the original image in both the horizontal and vertical directions. Increasing the stride can increase the speed but will result in some loss of image detail. A reasonable stride is beneficial to model construction and increasing the recognition speed of forest pests and diseases. Therefore, a stride experiment was designed for model design. Different strides correspond to precision and recall. To determine the optimal stride, the experiment was designed as follows: Figure 9 As shown.

[0162] 3. Add a detection layer

[0163] YOLOv5 is already lightweight enough, but feature extraction in the dataset detection part needs improvement. By incorporating the SElayer module, the model's sensitivity to features is enhanced, focusing on the relationships between channels. The model can automatically learn the importance of features in different channels. A Squeeze-and-Excitation (SE) module is proposed, such as... Figure 10 As shown, a small increase in computation yields a significant improvement in accuracy.

[0164] 4. Weighted frame fusion

[0165] Current object detection methods mostly use Non-Maximum Suppression (NMS) to select predicted boxes, but NMS can only retain one inaccurate box. Weighted Box Fusion (WBF) improves the accuracy of predicted boxes by utilizing information from all boxes. It can address situations where the model's predicted boxes are inaccurate. WBF uses information from three boxes to correct the predicted boxes. Methods are compared as follows... Figure 11 As shown, red represents predictions and blue represents true values.

[0166] (5) Deployment and testing: Further deploy and test in real environments, densely deploy heterogeneous wireless self-organizing networks with multiple sensor nodes in wetlands and woodlands and conduct autonomous test flights, maintain continuous operation for more than 6 months, collect and analyze data comprehensively, and optimize models and protocols based on the results.

[0167] In practical applications, the goal of navigation activities is to locate as many individual environmental data points as possible within a grid domain. Once the agent (i.e., the drone) occupies the same grid position as the target, the target is considered "located." To this end, this invention proposes a two-stream deep network architecture based on a map-based unified deep learning framework. This method computes grid-based navigation decisions {N, W, S, E} using two independent streams within a single deep inference network, based on real-time perception (tactical / exploitation) and historical navigation (strategic / exploration) information. This strategy can significantly outperform simpler strategies, such as the "lawnmower" pattern and other patterns.

[0168] This method uses a basic AlexNet design to process sensory input via a first-stream operation. The second-stream operates on the exploration history stored so far in a long-term memory map M. This stores the agent's current and past positions, as well as the location of diseased plants within a 20×20 grid of the flight area. The agent's starting position is fixed, and M is reset after δ% of the exploration map. These two streams are connected into a shallow integral network that maps to a SoftMaxNormalized likelihood vector V of possible navigation actions. During inference, the network selects the top-ranked navigation action from {N,W,S,E} based on V, executes it, and then updates the position history M. Figure 12This illustrates a real-world environment exploration example during one of the test flights of this invention.

[0169] This section presents eight examples of autonomous aircraft paths of varying lengths (due to different battery charging states) selected by the dual-stream navigation network for rapid target detection. The description shows a 20×20 exploration grid with the agent sensing local 5×5 cells on a 720×720 pixel image at a height of 10 meters. The start (orange) and end (green) grid cells of the experiment are also marked. Note that the determinism of the exploration agent architecture is visible even when no target is detected; the exploration in the bottom right corner (Experiment ID=18) illustrates this baseline motion pattern.

[0170] The repositioning command from {N,W,S,E} to the UAV flight platform needs to be issued via a local position offset, in meters, relative to the programmed Northeast (ENS) reference coordinate system. Therefore, to realize the target GPS coordinates generated by the exploration agent, they must be converted to this frame. This is achieved by converting the target GPS coordinates to a static geocentric fixed (ECEF) reference coordinate system, and then converting that coordinate to the local coordinate system. Similarly, the same process is performed on the agent's current GPS position, and the generated local position is compared.

[0171] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-source data-driven method for ecological protection and restoration, characterized in that, Includes the following steps: Step 1: Construct a multi-source data-driven wireless sensor network monitoring system, integrating wireless sensor networks and Mesh networks to form a heterogeneous wireless self-organizing network. Deploy multiple types of sensors to collect fine-grained ecological environment information of wetlands and forests, while simultaneously acquiring remote sensing image data via a UAV equipped with a binocular depth camera. Step 2: Based on the data collected by the wireless sensor network monitoring system, design an environmentally sensitive wetland and forest health and wellness index calculation model. Standardize the data, calculate climate comfort indicators, and perform principal component analysis to ultimately calculate and classify the health and wellness index. Step 3: Integrate the environmental data and remote sensing image data to form a time series. Construct a pine wilt disease prediction model based on R / S analysis and Cauchy kernel correlation vector machine. Combine UAV AI recognition and geographic information system to achieve precise location of diseased pine trees and epidemic prediction, thus realizing the prediction of pine wilt disease and precise location of diseased plants. Step 4: Optimize the MAC protocol, routing protocol, and transmission control protocol of the wireless sensor network, integrate environmental perception data and node location information to achieve cross-layer collaboration, and build a heterogeneous wireless ad hoc network that integrates the wireless sensor network and the Mesh network to ensure the efficiency and reliability of data transmission; Step 5: Deploy the heterogeneous wireless ad hoc network in wetlands and woodlands and conduct long-term testing for at least six months. Optimize the model and protocol based on the test results.

2. The multi-source data-driven ecological protection and restoration method according to claim 1, characterized in that: The multiple types of sensors mentioned in step one include at least two of the following: wind speed sensor, air pressure sensor, light sensor, sound wave sensor, humidity sensor, temperature sensor, raindrop sensor, negative ion sensor, PM2.5 sensor, noise sensor, carbon dioxide sensor, and ultraviolet radiation sensor; the drone is a Phase One X450 drone, which is equipped with an Intel d435i binocular depth camera and an NVIDIA Jetson TX2 embedded processing platform; the embedded processing platform is used to pre-train a YOLOv5 target detection model to identify pine tree diseases and pests in remote sensing images. The YOLOv5 model improves detection accuracy by adjusting the anchor structure, optimizing the step size, adding detection layers, and using a weighted bounding box fusion algorithm.

3. The multi-source data-driven ecological protection and restoration method according to claim 1, characterized in that: The construction process of the health and wellness index calculation model described in step two includes: using the formula Calculate the climate comfort index S, where T is air temperature, RH is relative humidity, and V is wind speed. Standardize the indicators for negative ions, PM2.5, noise, carbon dioxide, ultraviolet radiation index, and climate comfort index. For indicators where a higher value is better, use the formula... Standardization is performed; for indicators where smaller is better, a formula is used. Standardization is carried out, among which Principal component analysis was performed, and F1 = comfort * 0.826 - UV * 0.78 + negative ions * 0.739 - noise * 0.

721. F2 = PM2.5 * 0.924 F3 = carbon dioxide * 0.948 All parameters are standardized values. The Forest Health and Wellness Index (FHCI) is calculated as follows: F1×0.519255438908815+F2×0.260347671283085+F3×0.2203968898081. Based on the FHCI value, the health and wellness index is divided into four levels: FHCI≥0.60 is level 1, 0.20≤FHCI<0.60 is level 2, 0.0≤FHCI<0.20 is level 3, and FHCI<0 is level 4.

4. The multi-source data-driven ecological protection and restoration method according to claim 1, characterized in that: The construction process of the pine wilt disease prediction model in step three includes: (1) calculating the Hurst exponent H of pine tree diameter growth data using the R / S analysis method, and analyzing the temporal characteristics of pine tree diameter growth; (2) constructing a regression function based on the Cauchy kernel correlation vector machine, using the Hurst exponent H temporal data as training samples, and establishing a pine wilt disease prediction model, wherein the Cauchy kernel function is... , where σ is the Cauchy kernel parameter.

5. The multi-source data-driven ecological protection and restoration method according to claim 4, characterized in that, The process of calculating the Hearst exponent H using the R / S analysis method includes: segmenting the time series to obtain multiple sub-intervals. ; Calculate each subinterval Cumulative deviation : in, Reflects the cumulative shift of data points relative to the mean. This is time series data. sub-interval The mean; calculate the mean of each subinterval. range and standard deviation Calculate the rescaled range ratio For different window lengths ,Establish and linear relationship Where H is the Hurst exponent and C is the intercept term; the training data in the time-series experimental data for measuring the diameter growth of pine trees. The training sample set is constructed as follows: ， In the formula: X is the input sample set, and Y is the corresponding output set; a Cauchy kernel function is constructed, and a regression function for a Cauchy kernel correlation vector machine for predicting the time series of pine tree diameter growth is established based on the training sample set: In the formula: Indicates weight, This represents a mapping function.

6. The multi-source data-driven ecological protection and restoration method according to claim 1, characterized in that: The MAC protocol of the wireless sensor network described in step one is an environment-sensitive protocol. It integrates the results of lightweight environmental perception and prediction models, optimizes and adjusts the MAC layer's channel contention and avoidance mechanisms, and dynamically adjusts the MAC protocol's sleep and synchronization strategies to adapt to the environmental characteristics of wetlands, woodlands, high humidity, and multiple interferences. The routing protocol is a reliable routing protocol that integrates cross-layer and geographical location information. It utilizes the cross-layer interface extended by the protocol stack to obtain MAC layer information, thereby achieving efficient and reliable data transmission.

7. A multi-source data-driven ecological protection and restoration system, characterized in that, include: S1 Data Acquisition Module: Includes a multi-source sensor network and a UAV remote sensing unit. The multi-source sensor network is deployed in wetland and forest environments to collect fine-grained ecological and environmental information, and the UAV remote sensing unit is used to acquire remote sensing image data of wetlands and forests. S2 Data Transmission Module: A heterogeneous wireless ad hoc network integrating wireless sensor networks and Mesh networks, used to ensure reliable transmission of data collected by the data acquisition module. S3 Data Processing and Analysis Module: Deployed on an embedded platform, including a health and wellness index calculation unit and a pine wilt disease prediction unit. S4 Control and Optimization Module: Used to optimize and adjust the communication protocol of the data transmission module and the model of the data processing and analysis module.

8. The multi-source data-driven ecological protection and restoration system according to claim 7, characterized in that: The UAV remote sensing unit also includes a navigation module, which adopts a dual-stream deep network architecture, combines real-time sensing information and historical navigation data, and outputs navigation decisions through SoftMax normalization to achieve autonomous environment exploration and target positioning.

9. The multi-source data-driven ecological protection and restoration system according to claim 7, characterized in that: The data processing and analysis module uses an incremental learning method to dynamically adjust the parameters of the environmental perception model to adapt to the dynamic changes in wetland and forest environments. The system supports protocol performance pre-evaluation on the NS3 simulation platform and continuously optimizes the model and protocol through outdoor measured data.

10. The multi-source data-driven ecological protection and restoration system according to claim 7, characterized in that: It also includes a ground control module, which communicates bidirectionally with the data transmission module to receive data processing results and send control commands to the UAV remote sensing unit, enabling the UAV to navigate and explore autonomously.

Images