System and method for monitoring distribution of plant communities along highway

Abstract

The invention relates to the technical field of ecological environment monitoring, and particularly discloses a system and a method for monitoring distribution of plant communities along an expressway. An unmanned aerial vehicle carries a hyperspectral imager and a multi-parameter sensor to synchronously obtain vegetation spectral data and environmental parameters of soil salinity, water content, pH value, temperature and humidity; constructing a multi-dimensional data set with space-time matching; an improved end member extraction algorithm is combined with a soil salinity threshold triggering mechanism to realize self-adaptive updating of a dynamic end member library; an environmental constraint adversarial generative network is introduced, virtual end members conforming to ecological characteristics are generated and verified, and the coverage capability of an end member library is improved; performing physiological and ecological dual verification on the end members based on the knowledge graph to form an optimized end member set; and an environmental adaptability weight is introduced in the unmixing process, confidence evaluation and fuzzy logic analysis are combined, a graded credibility vegetation map is output, and accurate identification and spatial positioning of invasive plants and local vegetation are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological environment monitoring, and in particular to a system and method for monitoring plant community distribution along a highway. Background Art

[0002] With the rapid expansion of my country's expressway network, the ecological environment along these routes faces increasingly severe challenges, particularly from the increasing damage caused by invasive alien plants to native ecosystems. Traditional vegetation monitoring methods rely primarily on manual inspections or low- to medium-resolution remote sensing imagery, making it difficult to achieve large-scale, high-precision, and real-time plant community identification and spatial distribution analysis. In recent years, hyperspectral imaging technology has demonstrated great potential in vegetation classification and ecological monitoring due to its ability to provide rich spectral information about land objects. Combined with the mobility advantages of unmanned aerial vehicle platforms, hyperspectral remote sensing has been widely used in agriculture, forestry, and ecological and environmental monitoring. Furthermore, with the development of artificial intelligence and big data processing technologies, quantitative analysis methods based on endmember extraction and spectral unmixing have gradually become an important means of extracting vegetation cover information.

[0003] The existing technology has the following deficiencies: Existing technologies generally lack the ability to generate virtual endmembers and mechanisms for verifying ecological plausibility, making them difficult to handle missing or abnormal endmembers. This limits their effectiveness in areas along highways with high heterogeneity and complex vegetation types. Finally, most systems fail to establish effective confidence assessment systems, resulting in a lack of interpretability and credibility grading in recognition results, making it difficult to provide strong support for ecological protection decision-making. Summary of the Invention

[0004] The purpose of the present invention is to provide a system and method for monitoring plant community distribution along a highway to solve the above-mentioned problems.

[0005] The purpose of the present invention can be achieved through the following technical solutions: A method for monitoring plant community distribution along a highway comprises the following steps: S1: Use the UAV hyperspectral imaging system to synchronously acquire vegetation spectral data and collect environmental parameters in real time to establish a multi-dimensional monitoring dataset with spatiotemporal matching; The environmental parameters include: soil salt content, water content, pH value, atmospheric temperature and atmospheric humidity; S2: Online endmember extraction is performed on vegetation spectral data acquired in real time. A dynamic endmember library with environmental adaptability is constructed by combining environmental parameters. The adaptive adjustment mechanism of endmember characteristics is triggered by setting a soil salinity threshold. S3: Construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and discriminator, and inject the verified virtual endmembers into the dynamic endmember library; S4: Verify the plant physiological characteristics of the dynamic endmember library and verify its ecological rationality based on the knowledge graph to form a double-verified optimized endmember set; S5: Constrained linear spectral unmixing is performed based on the optimized endmember set to generate vegetation distribution maps with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

[0006] As a further solution of the present invention, the online endmember extraction of vegetation spectral data acquired in real time specifically includes: The dynamic window segmentation technology is used to process the hyperspectral data acquired in real time into blocks, and each data window contains a preset number of pixels; Based on the improved vertex component analysis algorithm, preliminary endmember extraction is performed on the spectral data in each data window, including: Normalize the spectral data in each data window to eliminate the influence of light intensity differences; Construct an n-dimensional spectral feature space, where n is the number of bands, and calculate the covariance matrix of the data set; The fast projection pursuit algorithm is used to find the vertex with the maximum projection direction in the feature space as the candidate end member; Dynamically adjust the contribution weight of each band in endmember extraction according to the band signal-to-noise ratio; Through an iterative optimization process, it is ensured that the extracted end members satisfy the convex simplex vertex conditions; Perform spectral continuity test on the extracted end members, calculate the reflectance change rate of adjacent bands, set the change rate threshold, detect whether there are mutation points in the spectral curve, evaluate the smoothness index of the overall spectral curve, and mark and eliminate candidate end members that do not meet the continuity requirements; The purity of the extracted end members is verified by the spectral similarity matching algorithm, and the similarity threshold is set to screen qualified candidate end members; The verified end-members are stored in the temporary end-member library, and their spatiotemporal characteristic information is recorded.

[0007] As a further solution of the present invention: the construction of a dynamic end member library with environmental adaptability by combining environmental parameters comprises the following steps: Establish a mapping relationship matrix between end members and environmental parameters, and associate and store each end member with its corresponding soil salinity, water content, and pH; Set up a multi-level soil salinity threshold system, and automatically trigger the end-member feature recalibration mechanism when the real-time monitoring data exceeds the current threshold range; Based on the changing characteristics of environmental parameters, an adaptive weighted algorithm is used to dynamically adjust the endmember spectral parameters; The adjusted end members are re-verified, updated to the dynamic end member library and marked with environmental adaptability characteristics; The multi-level soil salinity threshold system includes three levels: low salinity threshold, medium salinity threshold and high salinity threshold.

[0008] As a further solution of the present invention: the construction of a generative adversarial network with environmental parameter constraints includes the following steps: Design a dual-channel generator architecture, where the first channel inputs a random noise vector and the second channel inputs an environmental parameter feature vector; An environmental parameter verification module is introduced into the discriminator to ensure that the generated spectrum matches the environmental conditions; A conditional adversarial loss function is used to constrain the correlation between the generated spectrum and the target environment parameters; Through a progressive training strategy, the basic spectral characteristics are optimized first, and then the environmental response characteristics are optimized; The design of the conditional adversarial loss function includes: Calculate the distance metric between the generated spectrum and the true spectral distribution; Evaluate the correlation coefficients between the generated spectra and environmental parameters; Combining spectral feature loss and environmental constraint loss to construct a composite objective function; Dynamically adjust the weight coefficient of each loss.

[0009] As a further solution of the present invention: the virtual end member verification and injection includes the following steps: Establish a virtual endmember quality assessment index system, including spectral continuity, physiological rationality, and environmental consistency; A multi-level screening mechanism is adopted to first retain qualified candidates through rapid pre-screening and then conduct detailed verification; Mark the generation parameters and environmental applicable conditions for the virtual end members that have passed the verification; Design a priority injection strategy and determine the update order of the end-element library based on the matching degree of current environmental parameters.

[0010] As a further solution of the present invention: the plant physiological characteristics verification includes the following steps: Establish a plant physiological characteristics verification model, which includes three verification dimensions: chlorophyll absorption characteristics, water absorption characteristics, and cell structure characteristics; For each candidate endmember in the dynamic endmember library, calculate its absorption depth and waveform characteristics in the characteristic band; Analyze the matching degree between the calculated results and the standard plant physiological parameter library; Mark and remove end members that do not conform to the physiological characteristics of the plant; The construction of the standard plant physiological parameter library includes: Collect spectral data of typical plants at different growth stages; Measure the corresponding physiological indicators of chlorophyll content and water content; Establish a quantitative relationship model between spectral characteristics and physiological parameters; Regularly updated with new plant species data.

[0011] As a further solution of the present invention: the ecological rationality verification based on the knowledge graph includes the following steps: Construct a knowledge map of plants and the environment, including species distribution patterns, habitat preferences, and community composition relationships; Extract the plant species and environmental parameters corresponding to the end members; Perform multi-hop reasoning in the knowledge graph to verify the rationality of the corresponding plant's appearance under the current environmental conditions; Modify or eliminate end members whose ecological rationality scores are below the threshold; The specific implementation of the multi-hop reasoning includes: Search the knowledge graph for typical habitat conditions of the corresponding plants; Analyze the degree of match between current environmental parameters and typical habitats; Determine the symbiotic relationship between the corresponding plant and other surrounding plants.

[0012] As a further solution of the present invention: the constrained linear spectral unmixing comprises the following steps: Construct an endmember abundance inversion model with constraints, setting abundance non-negativity constraints and normalization constraints; The end-member environmental adaptability weight is introduced to adjust the contribution weight of each end-member in the unmixing according to the matching degree of environmental parameters; An iterative optimization algorithm is used to solve the abundance distribution, and the constraints are dynamically adjusted in each iteration; Perform spatiotemporal feature analysis on the unmixing residuals to identify and correct abnormal unmixing results; The calculation of the end member environmental adaptability weight includes: Analyze the similarity between the current environmental parameters and the best adaptation environment of each end member; Establish a nonlinear mapping relationship between environmental parameter differences and weight adjustments; Weighting of end members under extreme environmental conditions; Update weight parameters in real time to adapt to environmental changes.

[0013] As a further solution of the present invention: the confidence assessment and vegetation distribution map generation includes the following steps: Design multi-dimensional confidence evaluation indicators, including the size of unmixing residuals, end-member matching, and environmental consistency; Fusion of the evaluation results of each dimension based on fuzzy logic method to generate a comprehensive confidence score; Adaptive threshold method is used to classify confidence scores into high, medium and low confidence areas; Visually display the recognition results of different confidence levels in the vegetation distribution map; The implementation of the fuzzy logic method includes: Define the membership function of each evaluation indicator; Design fuzzy inference rule base; The centroid method is used for defuzzification; Verify the reliability of fuzzy reasoning results.

[0014] A system for monitoring plant community distribution along a highway, comprising: A multi-source data collaborative acquisition module, which uses a drone hyperspectral imaging system to synchronously acquire vegetation spectral data and environmental parameters in real time to establish a multi-dimensional monitoring data set with spatiotemporal matching; A dynamic endmember modeling module, which performs online endmember extraction on vegetation spectral data acquired in real time and constructs a dynamic endmember library with environmental adaptability based on environmental parameters. The module triggers an adaptive adjustment mechanism for endmember characteristics by setting a soil salinity threshold. A virtual endmember generation module is used to construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and the discriminator, and inject the verified virtual endmembers into the dynamic endmember library; An endmember verification module, which verifies the plant physiological characteristics of the dynamic endmember library and verifies the ecological rationality based on the knowledge graph to form an optimized endmember set that has undergone dual verification; An intelligent unmixing output module performs constrained linear spectral unmixing based on an optimized endmember set to generate a vegetation distribution map with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

[0015] Beneficial effects of the present invention: (1) This paper constructs an efficient and intelligent monitoring system for plant community distribution along highways by deeply integrating UAV hyperspectral imaging technology with a multi-parameter environmental sensing system. This method realizes the synchronous acquisition of vegetation spectral information and key environmental factors such as soil salinity, water content, pH value, atmospheric temperature and humidity, and forms a structured and scalable multidimensional monitoring data set through high-precision spatiotemporal registration technology, providing a high-quality data foundation for subsequent analysis. In the endmember extraction stage, an improved vertex component analysis algorithm is combined with dynamic window segmentation, band weighted optimization and iterative constraint strategy to improve the robustness and real-time performance of endmember identification; at the same time, a generative adversarial network (CGAN) is introduced to generate virtual endmembers that conform to plant physiological characteristics based on environmental parameter constraints, and inject them into the dynamic endmember library through spectral continuity test, ecological rationality assessment and multi-level screening mechanism, effectively expanding the endmember coverage and enhancing the system's adaptability. On this basis, a dual verification mechanism integrating knowledge graph is constructed to deeply screen candidate endmembers from the two dimensions of plant physiological characteristics and ecological rationality to ensure that the selected endmembers are both biologically consistent and ecologically logical. Finally, based on the optimized endmember set, a linear spectral unmixing model with non-negativity and normalization constraints was established, and an environmental adaptability weight mechanism was introduced to dynamically adjust the contribution of each endmember based on differences in environmental parameters, significantly improving the unmixing accuracy in complex ecological environments. At the same time, a multidimensional confidence assessment system was designed that integrates unmixing residuals, endmember matching, and environmental consistency indicators. Combined with fuzzy logic reasoning methods, comprehensive scoring and credibility grading visualization were achieved, further enhancing the interpretability and application value of the results. This method not only achieves high-precision identification and spatial positioning of invasive plants and native vegetation, but also provides strong technical support and data assurance for dynamic monitoring, risk warning, and scientific management of ecosystems along highways.

[0016] (2) The present invention innovatively introduces a trigger mechanism based on soil salinity thresholds, constructs a technical framework for adaptive adjustment of end-member features, and realizes dynamic modeling and optimization updates of vegetation spectral end-members in different ecological environments. By setting a multi-level soil salinity threshold system (low, medium, and high), the system can perceive environmental changes in real time and automatically start the end-member feature recalibration process accordingly to ensure that the end-members used always match the plant spectral response law under the current habitat conditions. On this basis, the adversarial generative network (CGAN) architecture with environmental parameter constraints is further integrated, and a dual-channel input method is used to introduce random noise vectors and environmental feature vectors respectively. Combined with the conditional adversarial loss function and progressive training strategy, the efficient generation of virtual end-member spectra that meet the physiological characteristics of plants in specific ecological niches is achieved. This mechanism not only effectively compensates for the limited coverage and insufficient adaptability of the traditional end-member library, but also significantly enhances the system's generalization ability to complex ecological environments. Furthermore, during the spectral unmixing process, the present invention introduces an end-member environmental adaptability weighting mechanism. Based on the similarity between the current pixel's environmental parameters (such as salinity and humidity) and the optimal environment of each end-member, a nonlinear mapping relationship is established, dynamically adjusting the weight of each end-member's contribution in the unmixing model. This design, which deeply embeds environmental factors into the unmixing process, gives the system greater terrain adaptability and greater accuracy in identifying mixed vegetation types. It significantly improves the response speed and recognition sensitivity to environmental changes or marginal habitats, providing solid technical support and decision-making basis for the early detection of potential invasive species along highways, ecological risk warning, and precise management. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a method for monitoring plant community distribution along a highway according to the present invention; Figure 2 This is a flow chart of a system for monitoring plant community distribution along highways in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] See also Figure 1 As shown, the present invention is a method for monitoring plant community distribution along a highway, comprising the following steps: S1: Use the UAV hyperspectral imaging system to synchronously acquire vegetation spectral data and collect environmental parameters in real time to establish a multi-dimensional monitoring dataset with spatiotemporal matching; The environmental parameters include: soil salt content, water content, pH value, atmospheric temperature and atmospheric humidity; S2: Online endmember extraction is performed on vegetation spectral data acquired in real time. A dynamic endmember library with environmental adaptability is constructed by combining it with environmental parameters. The adaptive adjustment mechanism of endmember characteristics is triggered by setting a soil salinity threshold. S3: Construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and discriminator, and inject the verified virtual endmembers into the dynamic endmember library; S4: Verify the plant physiological characteristics of the dynamic endmember library and verify its ecological rationality based on the knowledge graph to form a double-verified optimized endmember set; S5: Constrained linear spectral unmixing is performed based on the optimized endmember set to generate vegetation distribution maps with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

[0021] In S1, vegetation spectral data is acquired synchronously through the UAV hyperspectral imaging system, and environmental parameters are collected in real time to establish a multi-dimensional monitoring data set with temporal and spatial matching. Specifically, it includes: Specifically, the drone is equipped with a hyperspectral imager, a multi-parameter environmental sensor array, and a GPS / IMU positioning module. The hyperspectral imager continuously scans ground vegetation in the visible to near-infrared wavelength range (e.g., 400-2500nm), acquiring high-resolution vegetation reflectance spectral images. The multi-parameter environmental sensor array includes soil salinity sensors (such as conductivity probes), soil moisture sensors (such as time-domain reflectometry (TDR) sensors), pH sensors, and atmospheric temperature and humidity sensors, all used to measure key environmental factors within the flight area in real time. The GPS / IMU module accurately records the spatial position and attitude information of each frame of the spectral image, ensuring spatiotemporal consistency in subsequent data processing.

[0022] During execution, the drone cruises along a pre-set route along the highway, typically at an altitude between 80 and 150 meters to balance spatial resolution and coverage. Hyperspectral image acquisition is triggered at set intervals (e.g., 1 second) or flight distance (e.g., 5 meters), simultaneously recording the current environmental parameters. The collected raw data is organized according to unified timestamps and spatial coordinates, forming a multidimensional dataset containing spectral, environmental, and geographic information. This dataset serves as the input for subsequent endmember extraction, dynamic modeling, and unmixing analysis.

[0023] Furthermore, before data collection, the hyperspectral imager undergoes radiometric calibration and geometric correction to eliminate systematic errors. Simultaneously, each environmental sensor undergoes on-site calibration to ensure data accuracy. To enhance data quality, ground control points can be introduced to assist with post-processing registration, improving the spatiotemporal matching accuracy between data from different sources. This approach enables comprehensive perception and data fusion of vegetation and its ecological environment along the highway, providing a reliable data foundation for subsequent intelligent analysis.

[0024] In S2, online endmember extraction is performed on the real-time vegetation spectral data. A dynamic endmember library with environmental adaptability is constructed in combination with environmental parameters. The adaptive adjustment mechanism of endmember characteristics is triggered by setting the soil salinity threshold, which includes: The online endmember extraction process, based on an improved Vertex Component Analysis (VCA) algorithm, is suitable for the rapid processing of hyperspectral remote sensing imagery. First, the hyperspectral image data transmitted in real time by the drone is segmented using a dynamic window segmentation technique. This technique divides the entire image into multiple continuous data windows, each containing a preset number of pixels (e.g., 100×100 pixels). This reduces computational complexity and improves processing efficiency. Subsequent endmember extraction operations are performed independently on each data window, ensuring real-time performance and scalability.

[0025] Before performing initial endmember extraction, the spectral data within each data window undergoes normalization preprocessing to eliminate spectral deviations caused by variations in light intensity. Specifically, the spectral vector of each pixel is normalized to maintain consistency in energy or area, thereby improving the consistency of spectral data across time and regions.

[0026] The normalized spectral data is then constructed into an n-dimensional spectral feature space, where n is the number of bands in the hyperspectral image (e.g., 256 bands). Based on this, the covariance matrix of the entire dataset is calculated to describe the correlation between the bands and the data distribution characteristics.

[0027] Next, the Fast Projection Pursuit algorithm is used to search for extreme points in the direction of maximum projection in this feature space. These extreme points serve as candidate endmembers. This process simulates the vertex search of a convex hull and helps identify endmembers that best represent the extreme spectral characteristics in the dataset.

[0028] To further improve the accuracy of endmember extraction, the contribution weight of each band in endmember extraction is dynamically adjusted based on its signal-to-noise ratio (SNR). Specifically, bands with higher SNRs are given higher weights to enhance their influence in endmember identification; conversely, bands with lower SNRs are appropriately weighted to reduce noise interference.

[0029] Based on this, an iterative optimization process ensures that the extracted endmembers satisfy the convex simplex vertex condition, meaning that the spectra of all other pixels can be expressed as linear combinations of these endmembers. This process uses a non-negative matrix factorization strategy under a minimum volume constraint, continuously adjusting the endmember set until convergence.

[0030] To eliminate anomalous endmembers, the extracted results must be tested for spectral continuity. Specifically, the rate of change of reflectance in adjacent bands is calculated, and a reasonable threshold (e.g., ±0.05 / band) is set to detect the presence of abrupt changes. The smoothness of the overall spectral curve (e.g., the root mean square of the second derivative) is also evaluated, and candidate endmembers that do not meet continuity requirements are marked and eliminated.

[0031] In addition, spectral similarity matching algorithms (such as spectral angle matching (SAM), Euclidean distance (ED), or correlation coefficient (CC)) are used to verify the purity of the extracted endmembers. A similarity threshold (e.g., SAM < 0.1 radian) is set to screen out qualified endmembers that meet the criteria. Finally, qualified endmembers are stored in a temporary endmember library, along with their corresponding spatiotemporal coordinates, acquisition time, and original spectral data, for subsequent modeling.

[0032] After endmember extraction is complete, the dynamic endmember library construction phase begins. This involves establishing a mapping matrix between endmembers and environmental parameters. Each endmember record is associated with the environmental parameters corresponding to the moment of collection, such as soil salinity, moisture content, and pH, to form an "endmember-environment" attribute table. This structured storage method supports subsequent endmember recall and adaptive updates based on environmental conditions.

[0033] Furthermore, a multi-level soil salinity threshold system is established as a key control condition for triggering the adaptive adjustment mechanism of endmember characteristics. This multi-level threshold system includes three levels: low salinity threshold (e.g., EC < 4 dS / m), medium salinity threshold (e.g., 4 ≤ EC < 8 dS / m), and high salinity threshold (e.g., EC ≥ 8 dS / m), corresponding to different levels of suitability for plant growth. When the real-time monitored soil salinity data exceeds the threshold range of the current endmember, the endmember characteristic recalibration mechanism is automatically triggered to reevaluate and adjust the endmember's spectral characteristics.

[0034] Specifically, an adaptive weighted algorithm dynamically adjusts endmember spectral parameters based on the changing characteristics of environmental parameters. For example, if soil salinity in a region increases, the system will modify key characteristics of the original endmember, such as the absorption peak position and reflectance amplitude, based on historical data showing the spectral response of plants in high-salt environments. This will generate new endmembers that better reflect the current environmental conditions.

[0035] Finally, the adjusted endmembers undergo a secondary verification process, including spectral quality assessment and ecological plausibility checks. Once confirmed, they are updated to the dynamic endmember library and labeled with their applicable environmental adaptability characteristics (such as "salt-alkali tolerance" and "humid environment adaptability"). This dynamic endmember library supports incremental updates and version management to ensure that it always reflects the latest ecological environment and vegetation composition characteristics.

[0036] In summary, through the above-mentioned technical means, the present invention realizes the efficient extraction and environmental adaptive modeling of vegetation end members along highways, providing a solid data foundation and technical support for subsequent virtual end member generation, ecological verification and precise unmixing.

[0037] In S3, a generative adversarial network with environmental parameter constraints is constructed. Through collaborative training of the generator and the discriminator, virtual endmember spectra that conform to plant physiological characteristics are generated, and the verified virtual endmembers are injected into the dynamic endmember library. Specifically, The proposed GAN employs a dual-channel generator architecture. The first channel receives a random noise vector as input, introducing diversity into the generated spectrum. The second channel receives a set of environmental parameter feature vectors as conditional inputs, including but not limited to key factors such as soil salinity, moisture content, pH, atmospheric temperature, and humidity. Based on these two inputs, the generator generates a virtual endmember spectral vector that satisfies reasonable spectral characteristics under the current environmental conditions.

[0038] In the discriminator, the traditional GAN ​​architecture is extended to incorporate an environmental parameter verification module. The discriminator receives an input spectrum and the corresponding environmental parameters and outputs a probability indicating whether the spectrum originates from a real dataset. Furthermore, an auxiliary subnetwork is incorporated within the discriminator to assess the physical plausibility between the generated spectrum and the environmental parameters. For example, this subnetwork determines whether the generated spectrum exhibits the spectral response characteristics typical of salt-tolerant plants (e.g., shifted chlorophyll absorption peaks, deeper water absorption valleys), under high-salt conditions. This verification module can be implemented using a multilayer perceptron or a convolutional neural network, with its output participating in model training as part of the discriminant loss.

[0039] To enhance the correlation between the generated spectra and the target environmental parameters, this paper introduces a conditional adversarial loss function. This loss function consists of two core components: a spectral feature loss, which measures the similarity between the generated spectra and the real samples by calculating the distance between the spectral curves, such as the Wasserstein distance, KL divergence, or cosine similarity; and an environmental constraint loss, which evaluates the correlation between the generated spectra and the input environmental parameters. This can be quantified using the Pearson correlation coefficient, mutual information, or other statistical metrics.

[0040] The final objective function is the weighted sum of the two losses mentioned above. The weight of each loss can be dynamically adjusted according to the training stage. For example, in the early stages of training, optimization of basic spectral features is prioritized, and the weight of spectral feature losses is relatively high. As training progresses, the weight of environmental constraints is gradually increased to ensure that the generated results are more closely aligned with the plant's physiological performance under the target environmental conditions.

[0041] Furthermore, the present invention employs a progressive training strategy. Initially, the generator and discriminator are trained using standard vegetation spectral samples to develop basic spectral generation capabilities. Subsequently, different combinations of environmental parameters are gradually introduced to guide the model's learning of the effects of environmental factors on plant spectra. This strategy improves model stability and prevents model collapse caused by environmental interference during early training.

[0042] After the training of the generative adversarial network is completed, the virtual endmember verification and injection process begins. Specifically, a complete set of virtual endmember quality assessment indicators is established, including three dimensions: spectral continuity, physiological rationality, and environmental consistency: Spectral continuity assessment: Detect the smoothness of the generated spectral curve, calculate the reflectivity change rate of adjacent bands and set the threshold, and eliminate abnormal end members with mutation points; Physiological plausibility assessment: Based on a database of plant physiological parameters, verify whether the generated spectrum conforms to the typical absorption characteristics of known plant species, such as the chlorophyll a absorption peak at 680 nm and the water absorption valley at 970 nm; Environmental consistency assessment: Combined with current environmental parameters, determine whether the generated spectrum matches the common plant types in the habitat. For example, in a high-salt environment, the spectral characteristics should be closer to those of salt-tolerant species such as Suaeda salsa and Phragmites australis.

[0043] On this basis, a multi-stage screening mechanism is used to process the generated candidate virtual endmembers. The first stage is a rapid pre-screening, retaining only candidates that meet basic spectral continuity and physiological plausibility. The second stage is a detailed verification, combining knowledge graph reasoning, ecological adaptability models and other methods to comprehensively judge their applicability in specific environments.

[0044] For verified virtual endmembers, the system automatically labels their generation parameters (such as input noise, environmental parameters, and generation timestamp) and applicable environmental conditions (e.g., "adapted to saline-alkali soil" and "adapted to arid environments") and stores them in a temporary cache. Subsequently, a priority injection strategy is designed based on the matching degree between the current environmental state and each endmember in the historical endmember library. For example, if the soil salinity in the current area is high, virtual endmembers with a high degree of matching with high-salinity environments will be prioritized for injection into the dynamic endmember library to improve the accuracy and adaptability of subsequent unmixing analysis.

[0045] In summary, through the above technical solution, the present invention realizes the efficient generation and precise injection of virtual end members under complex environmental conditions, effectively expands the coverage and adaptability of the dynamic end member library, and thus improves the overall accuracy and robustness of plant community distribution monitoring along highways.

[0046] In S4, the dynamic endmember library is verified for plant physiological characteristics and ecological rationality based on the knowledge graph to form a double-verified optimized endmember set, including: The plant physiological characteristics verification process is achieved by establishing a specialized plant physiological characteristics verification model. This model incorporates three key verification dimensions: chlorophyll absorption characteristics, water absorption characteristics, and cellular structure characteristics. These three dimensions correspond to representative absorption peaks and reflectance platforms in the vegetation spectrum, effectively reflecting the health and physiological characteristics of the plant.

[0047] Specifically, for each candidate endmember in the dynamic endmember library, the system automatically calculates its absorption depth and waveform characteristics across multiple characteristic wavelengths. For example, it detects the absorption valley of chlorophyll a around 680nm, identifies the water absorption peak around 970nm, and analyzes the scattering characteristics caused by the internal structure of leaves between 1400nm and 1900nm. These characteristic parameters are then extracted and used in subsequent matching analysis.

[0048] The calculated results are then compared with a standard plant physiological parameter library. This library is constructed by collecting long-term hyperspectral data from typical plants at different growth stages (such as germination, growth, maturity, and senescence), and simultaneously measuring corresponding physiological indicators such as chlorophyll content and water content. Based on this, the system establishes a quantitative relationship model between spectral characteristics and physiological parameters, which can be used to evaluate whether candidate endmembers conform to the physiological performance of real plants.

[0049] If the spectral characteristics of a candidate endmember deviate significantly from the known plant physiological characteristics in the standard database, the system will mark it as an abnormal endmember and remove it from the dynamic endmember library to ensure that the endmember used in subsequent unmixing analysis is biologically accurate.

[0050] Furthermore, to further enhance the ecological suitability of endmembers, this paper introduces an ecological plausibility verification mechanism based on a knowledge graph. Specifically, the system constructs a plant-environment knowledge graph encompassing plant species, environmental factors, geographic distribution, habitat preferences, and interspecific relationships. This knowledge graph is implemented using graph database technology, with nodes representing entities (e.g., plant species, soil type, climatic conditions), and edges representing semantic relationships between entities (e.g., "adapted to," "symbiotic," and "commonly found in").

[0051] During the verification process, the system first extracts the plant species represented by the candidate endmembers and their corresponding environmental parameter information, such as soil salinity, humidity, and pH value. Then, it performs multi-hop reasoning in the knowledge graph to complete the following tasks in sequence: Find the typical habitat conditions of the plant species recorded in the knowledge graph; Analyze the degree of match between current environmental parameters and the typical habitat, for example, determine whether the soil salinity is within the range to which the species can adapt; Check the symbiotic relationship between the plant and the plants represented by other end members to confirm its rationality in the ecological community.

[0052] Through this reasoning process, the system generates an ecological plausibility score, which quantifies the credibility of a candidate endmember in the ecological context. If the score falls below a set plausibility threshold, the endmember is deemed less likely to occur in the current environment and the system will modify or directly remove it.

[0053] To enhance reasoning capabilities, the knowledge graph supports regular updates and expansions, adding new plant species, new environmental response patterns, and regional ecological survey results, thereby continuously improving the accuracy and coverage of ecological rationality verification.

[0054] In summary, through the dual verification mechanism combining plant physiological characteristics verification with ecological rationality verification, the present invention realizes the comprehensive screening and optimization of candidate endmembers in the dynamic endmember library, and finally forms an optimized endmember set that conforms to the laws of plant physiology and has ecological rationality, providing a solid data foundation and technical guarantee for subsequent high-precision vegetation distribution unmixing and invasive plant identification.

[0055] In S5, constrained linear spectral unmixing is performed based on an optimized endmember set to generate a vegetation distribution map with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation. Specifically, it includes: The constrained linear spectral unmixing process is based on an optimized endmember set that has been double-validated (for plant physiological characteristics and ecological rationality) to construct an endmember abundance inversion model with physical constraints. This model assumes that each pixel spectrum in the image is a linear combination of several endmember spectra, and satisfies two basic constraints: Non-negativity constraint: that is, the abundance value of each end member in the pixel must be greater than or equal to zero, reflecting that there is no "negative contribution" in reality; Normalization constraint: The sum of the abundances of all endmembers should be equal to 1, indicating that the pixel is composed entirely of these endmembers without additional unmodeled components.

[0056] On this basis, the system introduces an endmember environmental adaptability weighting mechanism to enhance the ecological applicability and regional adaptability of the unmixing results. Specifically, based on the currently collected environmental parameters (such as soil salinity, moisture content, and pH), the system analyzes their similarity with the optimal environmental conditions corresponding to each endmember. For example, if the soil salinity in a pixel is high, the endmember corresponding to plants that thrive in high-salt environments will be given a higher weight.

[0057] To more accurately represent the impact of environmental differences on endmember contributions, the system establishes a nonlinear mapping between environmental parameter differences and weight adjustments. For example, an exponential decay function or a sigmoid function is used to describe the trend of endmember weights decreasing as the degree of environmental deviation increases. For endmembers in extreme environmental conditions (such as extremely high salinity or low moisture), the system also implements specialized weighting strategies, such as setting a minimum weight threshold or directly blocking inappropriate endmembers, to avoid mismatches.

[0058] Furthermore, to address the challenges posed by complex terrain, mixed vegetation types, and spectral variation, the system employs an iterative optimization algorithm to solve the abundance distribution. During each iteration, the system dynamically adjusts constraints, for example, updating endmember weights based on the results of the previous iteration and correcting the initial estimates of anomalous pixels, thereby gradually approaching the optimal solution. Simultaneously, the system analyzes the spatiotemporal characteristics of the unmixing residuals (i.e., the difference between the observed and reconstructed spectra) to identify and correct anomalous unmixing results, such as incorrectly assigned areas caused by cloud shadows, water interference, or sensor noise.

[0059] After spectral unmixing is completed, the vegetation distribution map generation phase begins. To this end, the system has designed a multi-dimensional confidence assessment index system to quantify the reliability of the unmixing results. Specifically, it includes the following three core assessment dimensions: Unmixing residual size: measures the difference between the reconstructed spectrum and the original spectrum. The smaller the difference, the higher the confidence level. Endmember matching: evaluates whether the endmembers involved in unmixing are highly consistent with the actual vegetation in terms of spectral morphology; Environmental consistency: Determine whether the unmixing results are consistent with the plant distribution patterns under current environmental conditions.

[0060] In order to comprehensively integrate the information from the above multiple evaluation dimensions, the system uses fuzzy logic methods to integrate multi-source information. The specific implementation steps are as follows: Define the membership function: Set a reasonable membership function for each evaluation indicator, convert the original value into a fuzzy membership between 0 and 1, and indicate the credibility of the indicator within a specific confidence interval; Design a fuzzy inference rule base: Based on expert knowledge and historical data analysis, establish a set of fuzzy inference rules covering a variety of typical scenarios, such as "If the unmixing residual is small and the endmember matching is high, then the confidence level is high"; Defuzzification is performed using the centroid method: fuzzy inference results are converted into specific confidence scores to facilitate subsequent classification and visualization; Verify the reliability of fuzzy inference results: ensure the output of the fuzzy inference system is stable and reasonable through cross-validation, artificial sample testing, etc.

[0061] Finally, the system uses an adaptive threshold method to classify the comprehensive confidence score into high-confidence, medium-confidence, and low-confidence areas. High-confidence areas indicate high accuracy in vegetation identification and can be used for automatic classification and statistics. Medium-confidence areas indicate uncertainty and recommend manual review. Low-confidence areas may present a risk of misidentification and require specific annotation and on-site verification.

[0062] In visualizing vegetation distribution maps, the system color-codes identification results based on confidence levels, with high-confidence areas indicated in green, medium-confidence areas in yellow, and low-confidence areas in red. It also supports overlaying geographic information system (GIS) basemaps to enable spatial location and distribution trend analysis of invasive plants and native vegetation.

[0063] In summary, through the above technical scheme, the present invention realizes high-precision spectral unmixing, dynamic environmental adaptability modeling, multidimensional confidence assessment and visual map generation based on optimized end-member sets, which significantly improves the intelligence level and application practicality of plant community distribution monitoring along highways, and is particularly suitable for scenarios of early identification of invasive plants and ecological risk warning.

[0064] See also Figure 2 As shown, a system for monitoring plant community distribution along a highway includes: A multi-source data collaborative acquisition module, which uses a drone hyperspectral imaging system to synchronously acquire vegetation spectral data and environmental parameters in real time to establish a multi-dimensional monitoring data set with spatiotemporal matching; A dynamic endmember modeling module, which performs online endmember extraction on vegetation spectral data acquired in real time and constructs a dynamic endmember library with environmental adaptability based on environmental parameters. The module triggers an adaptive adjustment mechanism for endmember characteristics by setting a soil salinity threshold. A virtual endmember generation module is used to construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and the discriminator, and inject the verified virtual endmembers into the dynamic endmember library; An endmember verification module, which verifies the plant physiological characteristics of the dynamic endmember library and verifies the ecological rationality based on the knowledge graph to form an optimized endmember set that has undergone dual verification; An intelligent unmixing output module performs constrained linear spectral unmixing based on an optimized endmember set to generate a vegetation distribution map with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

[0065] The working principle of the present invention is as follows: through the hyperspectral imaging system and multi-parameter environmental sensor array carried on the UAV, the spectral data of vegetation and environmental parameters such as soil salinity, water content, pH value, atmospheric temperature and humidity are synchronously acquired, and a multi-dimensional monitoring data set with spatiotemporal matching is established in combination with the GPS / IMU module; based on the improved vertex component analysis algorithm, the real-time acquired hyperspectral images are subjected to online endmember extraction, and the efficiency and accuracy of endmember extraction are improved by adopting dynamic window segmentation, fast projection pursuit, band weighted optimization, iterative constraint solving and other methods, and a dynamic endmember library with environmental adaptability is constructed in combination with environmental parameters, in which a multi-level soil salinity threshold system is set to trigger the adaptive adjustment mechanism of endmember characteristics; a generative adversarial network (CGAN) with environmental parameter constraints is constructed, a dual-channel generator structure is used to input random noise vectors and environmental parameter vectors respectively, an environmental verification module is introduced into the discriminator, a conditional adversarial loss function including spectral feature loss and environmental constraint loss is designed, virtual endmember spectra that conform to plant physiological characteristics are generated through a progressive training strategy, and are injected into the dynamic endmember library after quality assessment; The candidate endmembers in the endmember database are double-validated. On the one hand, the plant physiological characteristic verification model is used to test whether the chlorophyll absorption peaks, water absorption valleys, and cell structure characteristics are consistent with the characteristics of real plants. On the other hand, the ecological rationality is verified based on the constructed plant-environment knowledge graph. Multi-hop reasoning is used to determine the habitat adaptability and interspecific symbiotic relationships of the plants represented by the endmembers in the current environment. Anomalous endmembers that do not conform to ecological logic are eliminated to form an optimized endmember set. Based on the optimized endmember set, a linear spectral unmixing model with non-negativity and normalization constraints is constructed. An endmember environmental adaptability weight mechanism is introduced to dynamically adjust the contribution of each endmember in the unmixing process according to the differences in environmental parameters. An iterative optimization algorithm is used to solve the abundance distribution. Anomalous results are corrected by combining the spatiotemporal characteristics of the unmixing residuals to finally generate a vegetation distribution map. A multi-dimensional confidence assessment system is designed, which integrates the size of the unmixing residuals, the endmember matching degree, and the environmental consistency index. A fuzzy logic method is used to calculate the comprehensive score, and an adaptive threshold method is used to divide the confidence levels into high, medium, and low areas. These areas are visually annotated in the map to provide a basis for subsequent manual review or automatic decision-making.

[0066] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for monitoring plant community distribution along a highway, characterized in that: The following steps are involved: S1: Use the UAV hyperspectral imaging system to synchronously acquire vegetation spectral data and collect environmental parameters in real time to establish a multi-dimensional monitoring dataset with spatiotemporal matching; The environmental parameters include: soil salt content, water content, pH value, atmospheric temperature and atmospheric humidity; S2: Online endmember extraction is performed on vegetation spectral data acquired in real time. A dynamic endmember library with environmental adaptability is constructed by combining environmental parameters. The adaptive adjustment mechanism of endmember characteristics is triggered by setting a soil salinity threshold. S3: Construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and discriminator, and inject the verified virtual endmembers into the dynamic endmember library; S4: Verify the plant physiological characteristics of the dynamic endmember library and verify its ecological rationality based on the knowledge graph to form a double-verified optimized endmember set; S5: Constrained linear spectral unmixing is performed based on the optimized endmember set to generate vegetation distribution maps with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

2. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The online endmember extraction of vegetation spectral data acquired in real time specifically includes: The dynamic window segmentation technology is used to process the hyperspectral data acquired in real time into blocks, and each data window contains a preset number of pixels; Based on the improved vertex component analysis algorithm, preliminary endmember extraction is performed on the spectral data in each data window, including: Normalize the spectral data in each data window to eliminate the influence of light intensity differences; Construct an n-dimensional spectral feature space, where n is the number of bands, and calculate the covariance matrix of the data set; The fast projection pursuit algorithm is used to find the vertex with the maximum projection direction in the feature space as the candidate end member; Dynamically adjust the contribution weight of each band in endmember extraction according to the band signal-to-noise ratio; Through an iterative optimization process, it is ensured that the extracted end members satisfy the convex simplex vertex conditions; Perform spectral continuity test on the extracted end members, calculate the reflectance change rate of adjacent bands, set the change rate threshold, detect whether there are mutation points in the spectral curve, evaluate the smoothness index of the overall spectral curve, and mark and eliminate candidate end members that do not meet the continuity requirements; The purity of the extracted end members is verified by the spectral similarity matching algorithm, and the similarity threshold is set to screen qualified candidate end members; The verified end-members are stored in the temporary end-member library, and their spatiotemporal characteristic information is recorded.

3. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The method of constructing a dynamic end member library with environmental adaptability by combining environmental parameters comprises the following steps: Establish a mapping relationship matrix between end members and environmental parameters, and associate and store each end member with its corresponding soil salinity, water content, and pH; Set up a multi-level soil salinity threshold system, and automatically trigger the end-member feature recalibration mechanism when the real-time monitoring data exceeds the current threshold range; Based on the changing characteristics of environmental parameters, an adaptive weighted algorithm is used to dynamically adjust the endmember spectral parameters; The adjusted end members are re-verified, updated to the dynamic end member library and marked with environmental adaptability characteristics; The multi-level soil salinity threshold system includes three levels: low salinity threshold, medium salinity threshold and high salinity threshold.

4. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The construction of the adversarial generative network with environmental parameter constraints includes the following steps: Design a dual-channel generator architecture, where the first channel inputs a random noise vector and the second channel inputs an environmental parameter feature vector; An environmental parameter verification module is introduced into the discriminator to ensure that the generated spectrum matches the environmental conditions; A conditional adversarial loss function is used to constrain the correlation between the generated spectrum and the target environment parameters; Through a progressive training strategy, the basic spectral characteristics are optimized first, and then the environmental response characteristics are optimized; The design of the conditional adversarial loss function includes: Calculate the distance metric between the generated spectrum and the true spectral distribution; Evaluate the correlation coefficients between the generated spectra and environmental parameters; Combining spectral feature loss and environmental constraint loss to construct a composite objective function; Dynamically adjust the weight coefficient of each loss.

5. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The virtual end member verification and injection comprises the following steps: Establish a virtual endmember quality assessment index system, including spectral continuity, physiological rationality, and environmental consistency; A multi-level screening mechanism is adopted to first retain qualified candidates through rapid pre-screening and then conduct detailed verification; Mark the generation parameters and environmental applicable conditions for the virtual end members that have passed the verification; Design a priority injection strategy and determine the update order of the end-element library based on the matching degree of current environmental parameters.

6. The method for monitoring plant community distribution along a highway according to claim 1, characterized in that: The plant physiological characteristics verification comprises the following steps: Establish a plant physiological characteristics verification model, which includes three verification dimensions: chlorophyll absorption characteristics, water absorption characteristics, and cell structure characteristics; For each candidate endmember in the dynamic endmember library, calculate its absorption depth and waveform characteristics in the characteristic band; Analyze the matching degree between the calculated results and the standard plant physiological parameter library; Mark and remove end members that do not conform to the physiological characteristics of the plant; The construction of the standard plant physiological parameter library includes: Collect spectral data of typical plants at different growth stages; Measure the corresponding physiological indicators of chlorophyll content and water content; Establish a quantitative relationship model between spectral characteristics and physiological parameters; Regularly updated with new plant species data.

7. The method for monitoring plant community distribution along a highway according to claim 1, characterized in that: The ecological rationality verification based on the knowledge graph includes the following steps: Construct a knowledge map of plants and the environment, including species distribution patterns, habitat preferences, and community composition relationships; Extract the plant species and environmental parameters corresponding to the end members; Perform multi-hop reasoning in the knowledge graph to verify the rationality of the corresponding plant's appearance under the current environmental conditions; Modify or eliminate end members whose ecological rationality scores are below the threshold; The specific implementation of the multi-hop reasoning includes: Search the knowledge graph for typical habitat conditions of the corresponding plants; Analyze the degree of match between current environmental parameters and typical habitats; Determine the symbiotic relationship between the corresponding plant and other surrounding plants.

8. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The constrained linear spectral unmixing comprises the following steps: Construct an endmember abundance inversion model with constraints, setting abundance non-negativity constraints and normalization constraints; The end-member environmental adaptability weight is introduced to adjust the contribution weight of each end-member in the unmixing according to the matching degree of environmental parameters; An iterative optimization algorithm is used to solve the abundance distribution, and the constraints are dynamically adjusted in each iteration; Perform spatiotemporal feature analysis on the unmixing residuals to identify and correct abnormal unmixing results; The calculation of the end member environmental adaptability weight includes: Analyze the similarity between the current environmental parameters and the best adaptation environment of each end member; Establish a nonlinear mapping relationship between environmental parameter differences and weight adjustments; Weighting of end members under extreme environmental conditions; Update weight parameters in real time to adapt to environmental changes.

9. The method for monitoring plant community distribution along a highway according to claim 1, wherein: The confidence assessment and vegetation distribution map generation includes the following steps: Design multi-dimensional confidence evaluation indicators, including the size of unmixing residuals, end-member matching, and environmental consistency; Fusion of the evaluation results of each dimension based on fuzzy logic method to generate a comprehensive confidence score; Adaptive threshold method is used to classify confidence scores into high, medium and low confidence areas; Visually display the recognition results of different confidence levels in the vegetation distribution map; The implementation of the fuzzy logic method includes: Define the membership function of each evaluation indicator; Design fuzzy inference rule base; The centroid method is used for defuzzification; Verify the reliability of fuzzy reasoning results.

10. A system for monitoring plant community distribution along a highway, characterized in that: A method for monitoring plant community distribution along a highway as claimed in any one of claims 1 to 9, comprising: A multi-source data collaborative acquisition module, which uses a drone hyperspectral imaging system to synchronously acquire vegetation spectral data and environmental parameters in real time to establish a multi-dimensional monitoring data set with spatiotemporal matching; A dynamic endmember modeling module, which performs online endmember extraction on vegetation spectral data acquired in real time and constructs a dynamic endmember library with environmental adaptability based on environmental parameters. The module triggers an adaptive adjustment mechanism for endmember characteristics by setting a soil salinity threshold. A virtual endmember generation module is used to construct a generative adversarial network with environmental parameter constraints, generate virtual endmember spectra that conform to plant physiological characteristics through collaborative training of the generator and the discriminator, and inject the verified virtual endmembers into the dynamic endmember library; An endmember verification module, which verifies the plant physiological characteristics of the dynamic endmember library and verifies the ecological rationality based on the knowledge graph to form an optimized endmember set that has undergone dual verification; An intelligent unmixing output module performs constrained linear spectral unmixing based on an optimized endmember set to generate a vegetation distribution map with confidence assessment, enabling accurate identification and spatial positioning of invasive plants and native vegetation.

Images