Vegetation type rapid identification method and system based on hyperspectral image

Abstract

The invention discloses a rapid vegetation type identification method and system based on a hyperspectral image, and relates to the field of vegetation type identification, and the method comprises the steps: obtaining hyperspectral image data and digital elevation data of a target region, and collecting time phenological information and climate region information of a geographic position; preprocessing the hyperspectral image data to obtain a surface reflectance data cube; extracting spectral features, spatial texture features and associated features of terrain and water systems of pixels in the cube, and constructing a multi-source feature vector; constructing a vegetation identification knowledge graph; constructing a space-time diagram convolutional network model, and outputting the probability of vegetation types in the target region and the overall classification confidence coefficient; and constructing a vegetation multi-level identification mechanism, and performing identification and drawing from canopy, sub-canopy / under-forest vegetation and ground cover vegetation. Efficient and automatic vegetation type exploration and identification can be performed on a large-range unknown area, and traditional low-efficiency and high-cost manual field investigation is replaced.

Description

Technical Field

[0001] This invention relates to the field of vegetation species identification, specifically to a method and system for rapid identification of vegetation species based on hyperspectral images. Background Technology

[0002] Vegetation is the main body of terrestrial ecosystems, and the spatial distribution and diversity of its species are core indicators reflecting the structure, function, and health of ecosystems. Obtaining vegetation species information quickly and accurately is of vital importance for agricultural and forestry management, biodiversity conservation, ecological environment monitoring, and global change research. Traditional ground vegetation surveys mainly rely on field investigations and manual identification. Although this method yields accurate results, it has inherent limitations such as low efficiency, high cost, strong subjectivity, and difficulty in covering large areas, and cannot meet the urgent need for "fast, accurate, and comprehensive" data in modern resource and environmental management.

[0003] The development of remote sensing technology, especially hyperspectral remote sensing, has provided a revolutionary technical approach for large-scale vegetation identification. Hyperspectral imaging technology can simultaneously acquire image information of ground objects in dozens to hundreds of continuous, narrow spectral bands, generating a continuous spectral curve for each pixel. This "image-spectrum integration" characteristic enables it to detect subtle differences in biochemical components such as pigments, cell structure, and water content of vegetation, thus providing far richer spectral information than traditional multispectral remote sensing for distinguishing different vegetation species. This unique spectral feature contained in the spectral curve is regarded as the "spectral fingerprint" of vegetation and is the physical basis for achieving precise vegetation identification.

[0004] In recent years, with the rapid development of machine learning, especially deep learning, significant progress has been made in the automatic vegetation classification technology based on hyperspectral images. Researchers have successively adopted traditional algorithms such as support vector machines and random forests, as well as deep learning models such as convolutional neural networks and recurrent neural networks to automatically learn and extract discriminative features from hyperspectral data, effectively improving classification accuracy. However, most existing methods remain within the single-plant recognition paradigm, that is, they identify plants by acquiring hyperspectral image data of a single plant or by identifying them by their appearance outline and color. This approach cannot quickly identify large areas, resulting in low efficiency in identifying vegetation in unknown target areas and wasting human and material resources. At the same time, existing deep learning models mainly focus on learning from the spectral and spatial information of the images themselves, without effectively incorporating prior knowledge of vegetation distribution and growth, ignoring the influence of the vegetation growth environment, and are easily misled by vegetation similar to those in hyperspectral image data, leading to recognition errors. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper provides a method and system for rapid vegetation species identification based on hyperspectral images. This technical solution solves the problem of the identification paradigm mentioned in the background technology, which is limited to the scope of a single plant and cannot quickly identify large areas. This results in low efficiency in identifying vegetation in unknown target areas, wasting human and material resources. At the same time, it fails to effectively incorporate prior knowledge of vegetation distribution and growth, ignores the influence of the vegetation growth environment, and is easily misled by vegetation with similar hyperspectral image data, leading to identification errors.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A rapid vegetation species identification method based on hyperspectral images includes: Acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area; Hyperspectral image data is preprocessed with radiometric calibration, atmospheric correction, and geometric correction to obtain a cube of surface reflectance data. Based on the surface reflectance data cube, the spectral features, spatial texture features, and correlation features with terrain and water system of its internal pixels are extracted to construct a multi-source feature vector; Based on multi-source features, a vegetation identification knowledge graph is constructed using the phenological observation network database and expert experience rules. Based on the vegetation recognition knowledge graph, a spatiotemporal graph convolutional network model is constructed to output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism was constructed to identify and map the canopy, sub-canopy / understory vegetation, and ground cover vegetation.

[0007] Preferably, the preprocessing of hyperspectral image data, including radiometric calibration, atmospheric correction, and geometric correction, to obtain a surface reflectance data cube specifically includes: Parse and obtain the metadata and raw digital quantization values ​​from the hyperspectral image data file, which is a standard format file containing a header file and a data body; Based on the radiometric calibration coefficients in the metadata, the original digital quantization values ​​are radiometrically calibrated and converted into physical apparent radiance data. An atmospheric correction method based on a physical radiative transfer model is used to perform atmospheric correction on apparent radiance data to eliminate atmospheric absorption and scattering effects and retrieve surface reflectance images. Ground control points are automatically selected using an image registration algorithm, and based on the coordinates of the control points, the coordinates of the control points are aligned and mapped with the geographic coordinates of the target area through affine transformation. After coordinate mapping, a pixel resampling method is used to fill the blank areas in the corrected image to obtain a geometrically corrected surface reflectance image. The corrected surface reflectance data for each band are filled into the data cube according to their spatial location and spectral wavelength to construct the surface reflectance data cube.

[0008] Preferably, the step of extracting the spectral features, spatial texture features, and correlation features with terrain and water systems of the internal pixels of the surface reflectance data cube to construct a multi-source feature vector specifically includes: Extract the continuous spectral reflectance curve of the target area pixel from the surface reflectance cube, and obtain its spectral characteristic parameters such as red edge position, green peak reflectance, and depth of water absorption valley. The gray-level co-occurrence matrix algorithm is used to calculate the spatial texture features in the neighborhood of the target region's pixels. The texture features include at least: homogeneity, contrast, and entropy. Calculate the Euclidean distance from the target area cell to the nearest water source as a water potential correlation feature, and obtain the elevation, slope, and aspect of its location from digital elevation data as terrain features. Spectral features, spatial texture features, water potential correlation features, and terrain features are spliced ​​and standardized to construct a multi-source feature vector.

[0009] Preferably, the step of constructing a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules specifically includes: Based on spectral feature parameters, spatial texture features, water potential correlation features, and terrain features, a node set of the knowledge graph is constructed. Based on statistical analysis and combined with expert experience rules, relationship edges between nodes are constructed. A knowledge graph for vegetation identification is constructed by using a graph database to store node sets and relational edges between nodes, where both nodes and relational edges contain quantitative attributes of their origin.

[0010] Furthermore, this solution proposes a rapid vegetation species identification system based on hyperspectral images to implement the aforementioned rapid vegetation species identification method based on hyperspectral images, including: The data acquisition module is used to acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area. The data processing module is used to preprocess hyperspectral image data by radiometric calibration, atmospheric correction, and geometric correction to obtain a surface reflectance data cube; based on the surface reflectance data cube, it extracts the spectral features, spatial texture features, and correlation features with topography and water system of its internal pixels to construct a multi-source feature vector; based on the multi-source features, and combined with the phenological observation network database and expert experience rules, it constructs a vegetation recognition knowledge graph. The vegetation identification module is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism is constructed to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively.

[0011] Preferably, the data processing module includes: The data preprocessing unit is used to preprocess the hyperspectral image data by radiometric calibration, atmospheric correction and geometric correction to obtain a surface reflectance data cube; The feature extraction unit is used to extract the spectral features, spatial texture features, and correlation features with terrain and water system of the internal pixels of the surface reflectance data cube, and construct a multi-source feature vector. The knowledge graph unit is used to construct a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules.

[0012] Preferably, the vegetation identification module includes: The vegetation identification unit is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. A multi-level identification unit is used to construct a multi-level vegetation identification mechanism based on the spatial distribution of vegetation, and to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method and system for rapid vegetation species identification based on hyperspectral images. This scheme obtains a clean and quantifiable surface reflectance data cube through a standardization process based on a physical radiative transfer model, involving radiometric calibration, atmospheric correction, and geometric correction. This eliminates sensor and atmospheric interference at the data source, laying a reliable physical data foundation for accurate identification. Furthermore, by fusing spectral features reflecting vegetation physiological states (red edge position, green peak reflectance, water absorption valley depth, etc.), spatial texture features characterizing canopy structure (homogeneity, contrast, entropy, etc.), and water potential and terrain correlation features depicting growth environment conditions, a multi-dimensional feature vector is constructed. This significantly enhances the ability to distinguish between "same spectral but different objects" and "same object with different spectral spectra" phenomena in complex scenes, avoiding the limitations of traditional methods that rely solely on spectral features. Based on this, by constructing a vegetation identification knowledge graph, multi-source features are deeply integrated with phenological patterns and expert experience to form a computable semantic network. This upgrades the vegetation identification process from simple data matching to intelligent reasoning driven by data and knowledge collaboration. It significantly improves the inherent defects of traditional deep learning models, such as weak generalization ability and uninterpretable decision-making when facing unknown species or growth environments, making the identification results more in line with ecological laws. On this basis, the multi-level vegetation identification mechanism proposed in this scheme, based on the spatial distribution of vegetation vertical structure, collaboratively utilizes the advantages of satellites, drones and ground robots to identify and map canopy, understory and ground cover vegetation respectively. It realizes a comprehensive and rapid survey from macro to micro, greatly improving the efficiency, accuracy and reliability of vegetation identification in large-scale unknown areas. Attached Figure Description

[0014] Figure 1 This is a flowchart of a method for rapid identification of vegetation species based on hyperspectral images according to the present invention. Figure 2 The flowchart of the preprocessing of hyperspectral image data for radiometric calibration, atmospheric correction and geometric correction to obtain surface reflectance data cube is shown below. Figure 3 The present invention extracts the spectral features, spatial texture features, and correlation features with terrain and water system of the internal pixels of the surface reflectance data cube, and constructs a multi-source feature vector flowchart. Figure 4 The flowchart for constructing a spatiotemporal graph convolutional network model for this invention outputs the probability of vegetation species in the target area and the overall classification confidence. Detailed Implementation

[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0016] Reference Figure 1As shown, a method for rapid identification of vegetation species based on hyperspectral images includes: Acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area; Hyperspectral image data is preprocessed with radiometric calibration, atmospheric correction, and geometric correction to obtain a cube of surface reflectance data. Based on the surface reflectance data cube, the spectral features, spatial texture features, and correlation features with terrain and water system of its internal pixels are extracted to construct a multi-source feature vector; Based on multi-source features, a vegetation identification knowledge graph is constructed using the phenological observation network database and expert experience rules. Based on the vegetation recognition knowledge graph, a spatiotemporal graph convolutional network model is constructed to output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism was constructed to identify and map the canopy, sub-canopy / understory vegetation, and ground cover vegetation.

[0017] This solution provides a technical approach that enables efficient and automated exploration and identification of vegetation species in large, unknown areas, replacing traditional, inefficient, and costly manual field surveys. Ultimately, it allows for the rapid and accurate mapping of vegetation species distribution in the target area, serving critical needs such as biodiversity surveys, the discovery of rare species, and baseline surveys of ecological resources.

[0018] The acquisition of airborne and spaceborne hyperspectral image data of the target area, and the simultaneous acquisition of digital elevation data, phenological information of the acquisition time, and climatic zone information of the geographical location of the area, specifically includes: Hyperspectral image data of the target area are acquired using airborne and spaceborne imaging spectrometers, with a spectral range covering the visible light to short-wave infrared band. Digital elevation data of the target area is obtained from geographic information platforms or field measurements, and its spatial resolution matches that of hyperspectral images. Record the specific date of hyperspectral image data acquisition and query the corresponding time phenological information from the phenological observation network database; Based on the geographical coordinates of the target area, determine its Köppen climate zone or similar climate classification, and extract the annual precipitation and annual mean temperature range of that climate zone as background information.

[0019] This can be explained by the fact that accurate identification of vegetation species cannot rely solely on spectral characteristics; their growth environment must also be considered. In the natural environment, there are often species with similar spectral characteristics but vastly different ecological habits. For example, both *Pinus tabuliformis* and *Pinus huangshanensis* are evergreen conifers, and their spectral reflectance curves in the visible-near-infrared band are highly similar, making them difficult to distinguish based on spectral analysis alone. However, their ecological distributions differ significantly: *Pinus tabuliformis* is mainly distributed in the temperate monsoon climate zone of northern China at altitudes of 100–2600 meters, with an annual precipitation of approximately 500–800 mm; while *Pinus huangshanensis* is distributed in the subtropical humid climate zone of southeastern China at altitudes of 600–1800 meters, with an annual precipitation of approximately 1200–2000 mm, and requires higher air humidity. Therefore, this scheme deeply integrates the target's growth environment information while analyzing hyperspectral image data. The system uses information to assist decision-making. Specifically, it constructs an integrated air-space-ground perception network using satellites, low-altitude drones, and land robots to acquire comprehensive data of the target area from different directions. The system simultaneously acquires phenological information at the time of collection (i.e., data, phenomena, or characteristics observed at a specific point in time that reflect the life cycle stages of organisms in nature and the rhythm of environmental changes, i.e., vegetation growth information and corresponding environmental change information) as well as key environmental backgrounds such as the Köppen climate zone of the target area. By using these environmental factors as constraints, the system can effectively eliminate ecologically unreasonable options during the identification process. For example, in the above case, if the target area is located in a subtropical mountainous area with high precipitation, the system will be more inclined to identify it as Huangshan pine, thereby achieving rapid and accurate identification and screening of spectrally similar tree species, greatly improving the accuracy and reliability of identification.

[0020] Reference Figure 2 As shown, the preprocessing of hyperspectral image data, including radiometric calibration, atmospheric correction, and geometric correction, to obtain a surface reflectance data cube specifically includes: Parse and obtain the metadata and raw digital quantization values ​​from the hyperspectral image data file, which is a standard format file containing a header file and a data body; Based on the radiometric calibration coefficients in the metadata, the original digital quantization values ​​are radiometrically calibrated and converted into physical apparent radiance data. An atmospheric correction method based on a physical radiative transfer model is used to perform atmospheric correction on apparent radiance data to eliminate atmospheric absorption and scattering effects and retrieve surface reflectance images. Ground control points are automatically selected using an image registration algorithm, and based on the coordinates of the control points, the coordinates of the control points are aligned and mapped with the geographic coordinates of the target area through affine transformation. After coordinate mapping, a pixel resampling method is used to fill the blank areas in the corrected image to obtain a geometrically corrected surface reflectance image. The corrected surface reflectance data for each band are filled into the data cube according to their spatial location and spectral wavelength to construct the surface reflectance data cube.

[0021] This can be explained by the fact that during the acquisition of hyperspectral images, the imaging spectrometer converts the received radiant energy into digital signals. These signals are recorded in the form of raw digital quantized values. However, these raw digital quantized values, i.e., DN values, are affected by a combination of factors such as sensor response characteristics, sunlight conditions, and atmospheric transport paths, and therefore lack direct physical comparability and cannot be directly used for detailed data analysis. Therefore, this scheme performs radiometric calibration based on the inherent radiometric calibration coefficients of the imaging spectrometer, converting the raw, relative DN values ​​into physically meaningful apparent radiance. An atmospheric correction method based on a physical radiative transfer model (such as using the FLAASH module based on the MODTRAN physical model in ENVI) is employed to simulate and eliminate the absorption and scattering effects of molecules such as water vapor and aerosols in the atmosphere, thereby retrieving the true intrinsic reflectance of the ground features. This is a crucial prerequisite for achieving high-precision vegetation spectral identification. Furthermore, a high-precision ground control point is selected through an image registration algorithm, and an affine transformation is used to establish a mapping relationship between the control point coordinates and the true geographic coordinates of the target area, effectively eliminating... To address image spatial distortion caused by factors such as sensor attitude and terrain undulation, the method of automatically selecting ground control points through image registration algorithms involves: registering the atmospherically corrected image with a geometrically corrected reference remote sensing map; automatically extracting and matching corresponding feature points on the two images using scale-invariant feature transformation or accelerated robust feature algorithms; identifying successfully matched feature points as ground control points; and filling blank areas in the corrected image using pixel resampling to avoid the impact of blank areas caused by geometric transformations after coordinate mapping. Finally, by integrating spatial and spectral reflectance data across all bands, a surface reflectance data cube is constructed, transforming the hyperspectral image from a simple "image" into a quantitative data set that is pure in the spectral dimension and accurate in the spatial dimension. Preprocessing effectively eliminates various systematic errors from data acquisition to transmission, enabling the data itself to be directly used for subsequent quantitative analysis and modeling, laying a solid data foundation for rapid and accurate identification of vegetation species.

[0022] Reference Figure 3 As shown, the step of extracting the spectral features, spatial texture features, and correlation features with terrain and water systems of the internal pixels of the surface reflectance data cube to construct a multi-source feature vector specifically includes: Extract the continuous spectral reflectance curve of the target area pixel from the surface reflectance cube, and obtain its spectral characteristic parameters such as red edge position, green peak reflectance, and depth of water absorption valley. The gray-level co-occurrence matrix algorithm is used to calculate the spatial texture features in the neighborhood of the target region's pixels. The texture features include at least: homogeneity, contrast, and entropy. Calculate the Euclidean distance from the target area cell to the nearest water source as a water potential correlation feature, and obtain the elevation, slope, and aspect of its location from digital elevation data as terrain features. Spectral features, spatial texture features, water potential correlation features, and terrain features are spliced ​​and standardized to construct a multi-source feature vector.

[0023] This solution constructs multi-source feature vectors to comprehensively characterize vegetation attributes from three dimensions: physiological, structural, and ecological. This enables precise identification. Spectral features (red edge position, green peak reflectance, and water absorption valley depth) directly reflect the intrinsic physiological states of vegetation, such as chlorophyll, nitrogen, and water content. Spatial texture features (homogeneity, contrast, and entropy) quantify the spatial distribution and complexity of the canopy structure. Topographic and water system association features provide key environmental backgrounds such as altitude, slope aspect, and distance to water sources, serving as constraints for determining the rationality of vegetation distribution. Through multi-source information fusion, the limitations of single spectral features in distinguishing between "same spectral features but different vegetation" or "same vegetation in different environments" are effectively addressed, significantly improving the accuracy and reliability of classification. The red edge position refers to a spectral curve within the wavelength range of 680 nm to 780 nm. The wavelength corresponding to the maximum value of the first derivative reflects the boundary of abrupt changes in chlorophyll content and growth vitality of vegetation. Green peak reflectance refers to the global maximum reflectance value and its corresponding reflectance value in the green visible light band of the spectral curve from 500 nm to 600 nm. This peak value reflects the strongest reflectance ability of vegetation to green light and is closely related to the chlorophyll and nitrogen status of leaves. Water absorption valley depth refers to the strong water absorption band near about 1150 nm, 1400 nm or 1900 nm. The parameter characterizes the water content of vegetation canopy by quantifying the degree of local concavity of the reflectance curve. A preferred implementation method is to quantify it by calculating the vertical distance between the reflectance at the bottom of the absorption valley and the reflectance of the two absorption shoulders (e.g., at 1050 nm and 1300 nm). This parameter directly characterizes the water content of vegetation canopy.

[0024] The construction of a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules specifically includes: Based on spectral feature parameters, spatial texture features, water potential correlation features, and terrain features, a node set of the knowledge graph is constructed. Based on statistical analysis and combined with expert experience rules, relationship edges between nodes are constructed. A knowledge graph for vegetation identification is constructed by using a graph database to store node sets and relational edges between nodes, where both nodes and relational edges contain quantitative attributes of their origin.

[0025] Explained by this, based on the acquisition of multi-source feature vectors, this scheme does not directly use the multi-source feature vectors as input. Instead, it constructs a vegetation identification knowledge graph to transform data from "statistical association" to "semantic reasoning." The vegetation identification knowledge graph explicitly transforms the implicit, discrete associations in the multi-source feature vectors (e.g., feature A often coexists with vegetation B) into structured, interpretable semantic relationships (e.g., species B "prefers" environment C). Furthermore, by embedding expert experience rules, it provides three capabilities for subsequent analysis: First, it uses paths in the graph for relational reasoning, allowing indirect judgment even if a vegetation species lacks spectral features, through its symbiotic species or preferred growth environment; second, it possesses expert experience rules, enabling logical judgment of whether a vegetation identification result is ecologically "reasonable"; and third, it achieves knowledge sharing and transfer, allowing new knowledge about a vegetation in the graph to be extended to other related vegetation species through the relational network. This deep understanding and application of domain knowledge cannot be achieved by simply inputting feature vectors. The node set includes: Vegetation species node: Known vegetation species recorded in the phenological observation network database; Spectral feature nodes: spectral feature parameters extracted from known vegetation species samples, including red edge position, green peak reflectance, and water absorption valley depth; Phenological nodes: Temporal phenological information for known vegetation species; Environmental factor nodes: Topographic features and water potential correlation features of known vegetation types; The relationship edges between the nodes specifically include: Calculate the similarity (e.g., spectral angular distance) of spectral feature nodes between any two vegetation type nodes. When the similarity is higher than the preset threshold S1 value, construct the spectral similarity edge between the two nodes. Based on expert experience rules, a preference relationship is set between vegetation type nodes and environmental factor nodes to construct an ecological preference edge. The preference relationship includes at least: suitable, unsuitable and tolerable. Based on the temporal phenological information in the phenological observation database, a phenological temporal association is set between vegetation type nodes and phenological period nodes, and phenological association edges are established. The phenological time series correlation includes at least: symbiosis (time synchronization), succession (time sequence) and mutual exclusion (time separation).

[0026] Reference Figure 4 As shown, the construction of the spatiotemporal graph convolutional network model, which outputs the probability of vegetation species in the target area and the overall classification confidence, specifically includes: The adjacency matrix and node feature matrix are extracted from the vegetation recognition knowledge graph and used as the input matrix of the graph neural network. The adjacency matrix is ​​extracted from the relation edges and the node feature matrix is ​​extracted from the node set. Based on known vegetation species samples in the phenological observation network database, true category labels are set; Based on graph convolutional networks and temporal convolutional networks, a spatiotemporal graph convolutional network model is constructed. With the goal of predicting the deviation between the class label and the true class label, a loss function is constructed by weighted summation of cross-entropy loss and L2 regularization loss; Based on known vegetation species samples in the phenological observation network database, a knowledge graph of all vegetation species is drawn, and the corresponding adjacency matrix and node feature matrix are extracted. Based on the adjacency matrix and node feature matrix, the FGSM method is used to introduce training adversarial sample data to construct a vegetation recognition training sample set; Based on the vegetation recognition training sample set, a node feature attention mechanism and an adjacency attention mechanism are constructed, and multiple attention weights are formed by weighted summation to obtain the fused features. The fused features are input into a Softmax classification layer, which outputs the probability of vegetation species in the target area and the overall classification confidence.

[0027] This can be explained by the fact that the node feature matrix reflects the vegetation characteristics of vegetation species, spectral features, phenological periods, and environmental factors, while the adjacency matrix reflects the similarity of spectral feature nodes between any two vegetation species nodes, the preference relationship between vegetation species nodes and environmental factor nodes, and the phenological time series association between vegetation species nodes and phenological period nodes. By inputting hyperspectral images collected from the target area and real-time collected phenological period and environmental factor data, the node feature matrix and adjacency matrix are obtained through the processing steps of this scheme. Then, the probability of vegetation species in the target area and the overall classification confidence are output through the spatiotemporal graph convolutional network model. The spectral features of vegetation are deeply integrated and analyzed with the growth environment, which effectively improves the accuracy and reliability of vegetation species identification. The overall classification confidence refers to the maximum probability value among all category probabilities output by the Softmax classification layer, which is used to measure the certainty of the model's classification result. The construction of the node feature attention mechanism and the adjacency attention mechanism specifically includes: Node feature attention: The multi-source feature vector of the pixel to be identified is matched with the node feature matrix of the knowledge graph, and the similarity between it and each node in the graph is calculated as the node feature attention weight. Environment and phenological attention: The real-time acquired phenological data is sequence matched with the features of phenological period nodes in the map, and their similarity is calculated as the dynamic phenological state attention weight; Environmental Constraint Attention: For each ecological preference edge in the graph, calculate the conformity (e.g., Gaussian kernel function value) between the real-time terrain features and water potential correlation features of the pixel to be identified and the typical environmental conditions (e.g., value range) of the environmental factor node pointed to by the edge, and aggregate the conformity of all edges as environmental constraint weights. It should be noted that the three attention weights are fused by weighted summation to form a multi-attention weight, and the weight update method can be learned by the fully connected layer from the weighted summation weights. Graph convolutional layers are used to aggregate features of adjacent nodes in a knowledge graph. The node feature update formula is as follows:

[0028] In the formula, For the first The node feature matrix of the layer, For the first The node feature matrix of the layer It is the ReLU activation function. It is a self-connected adjacency matrix. ,in, It is an adjacency matrix. It is the identity matrix. for The degree matrix, For the first The weight matrix of the graph convolution of the layers can be trained. For feature fusion operators, For temporal convolutional network operations, a one-dimensional dilated convolution kernel is used to perform convolution in the time dimension to extract the temporal relationships of spectral features, phenological periods, and environmental factors in the node feature matrix.

[0029] The aforementioned multi-level vegetation identification mechanism, based on the spatial distribution of vegetation, specifically includes the identification and mapping of canopy, sub-canopy / understory vegetation, and ground cover vegetation, and includes: Satellites, drones, and land robots serve as sensing terminals, responsible for acquiring hyperspectral image data and transmitting the acquired data to the cloud platform; The cloud platform integrates a vegetation recognition knowledge graph and a spatiotemporal graph convolutional network model to form a vegetation recognition intelligent processing center. A multi-level vegetation identification mechanism is constructed, in which the cloud platform collaboratively processes data from different sensing terminals to identify and map vegetation at each level of the vertical structure.

[0030] This can be explained by the fact that different vegetation layers exhibit significant differences in spectral characteristics and spatial scale, requiring targeted matching of sensing devices and identification methods to ensure accuracy. This solution constructs a multi-layered vegetation identification mechanism that integrates satellite, UAV, and land robot technologies. Based on the spatial resolution and scale differences of data from different sensing devices, it identifies canopy, sub-canopy / understory, and ground cover vegetation separately. This effectively overcomes the limitations of traditional single-scale identification, fully aligns with the ecological structure characteristics of vertical vegetation stratification, and achieves three-dimensional, all-round vegetation species identification from the macroscopic canopy to the microscopic ground cover. This provides more refined classification results for ecological monitoring, resource surveys, and other scenarios. Specifically, the multi-layered vegetation identification mechanism includes: Satellite identification layer: The cloud platform processes large-scale, low-to-medium spatial resolution data transmitted from satellites, and identifies and maps the dominant tree species in the canopy based on the spectral characteristics of macroscopic mixed pixels; Drone identification layer: The cloud platform processes high-resolution data of the target area transmitted back by drones, and identifies and maps the sub-canopy / understory vegetation based on the spectral characteristics at the scale of canopy or vegetation patches. Ground identification layer: The cloud platform processes high-resolution data of the target area transmitted back by the land robot, and identifies and maps the ground cover vegetation based on the spectral characteristics at the leaf or single plant scale. Through cloud-edge-device architecture and multi-layered vegetation identification mechanism, a three-dimensional and all-round vegetation species identification of the target area from the canopy to the ground cover layer is achieved.

[0031] Furthermore, based on the same inventive concept as the aforementioned rapid vegetation species identification method based on hyperspectral images, this solution proposes a rapid vegetation species identification system based on hyperspectral images, comprising: The data acquisition module is used to acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area. The data processing module is used to preprocess hyperspectral image data by radiometric calibration, atmospheric correction, and geometric correction to obtain a surface reflectance data cube; based on the surface reflectance data cube, it extracts the spectral features, spatial texture features, and correlation features with topography and water system of its internal pixels to construct a multi-source feature vector; based on the multi-source features, and combined with the phenological observation network database and expert experience rules, it constructs a vegetation recognition knowledge graph. The vegetation identification module is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism is constructed to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively. The data processing module includes: The data preprocessing unit is used to preprocess the hyperspectral image data by radiometric calibration, atmospheric correction and geometric correction to obtain a surface reflectance data cube; The feature extraction unit is used to extract the spectral features, spatial texture features, and correlation features with terrain and water system of the internal pixels of the surface reflectance data cube, and construct a multi-source feature vector. The knowledge graph unit is used to construct a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules. The vegetation identification module includes: The vegetation identification unit is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. A multi-level identification unit is used to construct a multi-level vegetation identification mechanism based on the spatial distribution of vegetation, and to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively.

[0032] In summary, the advantages of this invention are: it enables efficient and automated exploration and identification of vegetation species in large-scale unknown areas, replacing the traditional, inefficient, and costly manual field surveys.

[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for rapid identification of vegetation species based on hyperspectral images, characterized in that, include: Acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area; Hyperspectral image data is preprocessed with radiometric calibration, atmospheric correction, and geometric correction to obtain a cube of surface reflectance data. Based on the surface reflectance data cube, the spectral features, spatial texture features, and correlation features with terrain and water system of its internal pixels are extracted to construct a multi-source feature vector; Based on multi-source features, a vegetation identification knowledge graph is constructed using the phenological observation network database and expert experience rules. Based on the vegetation recognition knowledge graph, a spatiotemporal graph convolutional network model is constructed to output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism was constructed to identify and map the canopy, sub-canopy / understory vegetation, and ground cover vegetation.

2. The method for rapid identification of vegetation species based on hyperspectral images according to claim 1, characterized in that, The acquisition of airborne and spaceborne hyperspectral image data of the target area, and the simultaneous acquisition of digital elevation data, phenological information of the acquisition time, and climatic zone information of the geographical location of the area, specifically includes: Hyperspectral image data of the target area are acquired using airborne and spaceborne imaging spectrometers, with a spectral range covering the visible light to short-wave infrared band. Digital elevation data of the target area is obtained from geographic information platforms or field measurements, and its spatial resolution matches that of hyperspectral images. Record the specific date of hyperspectral image data acquisition and query the corresponding time phenological information from the phenological observation network database; Based on the geographical coordinates of the target area, determine its Köppen climate zone or similar climate classification, and extract the annual precipitation and annual mean temperature range of that climate zone as background information.

3. The method for rapid identification of vegetation species based on hyperspectral images according to claim 2, characterized in that, The preprocessing of hyperspectral image data, including radiometric calibration, atmospheric correction, and geometric correction, to obtain a surface reflectance data cube specifically includes: Parse and obtain the metadata and raw digital quantization values ​​from the hyperspectral image data file, which is a standard format file containing a header file and a data body; Based on the radiometric calibration coefficients in the metadata, the original digital quantization values ​​are radiometrically calibrated and converted into physical apparent radiance data. An atmospheric correction method based on a physical radiative transfer model is used to perform atmospheric correction on apparent radiance data to eliminate atmospheric absorption and scattering effects and retrieve surface reflectance images. Ground control points are automatically selected using an image registration algorithm, and based on the coordinates of the control points, the coordinates of the control points are aligned and mapped with the geographic coordinates of the target area through affine transformation. After coordinate mapping, a pixel resampling method is used to fill the blank areas in the corrected image to obtain a geometrically corrected surface reflectance image. The corrected surface reflectance data for each band are filled into the data cube according to their spatial location and spectral wavelength to construct the surface reflectance data cube.

4. The method for rapid identification of vegetation species based on hyperspectral images according to claim 3, characterized in that, The process of extracting spectral features, spatial texture features, and correlation features with terrain and water systems from the internal pixels of the surface reflectance data cube to construct a multi-source feature vector specifically includes: Extract the continuous spectral reflectance curve of the target area pixel from the surface reflectance cube, and obtain its spectral characteristic parameters such as red edge position, green peak reflectance, and depth of water absorption valley. The gray-level co-occurrence matrix algorithm is used to calculate the spatial texture features in the neighborhood of the target region's pixels. The texture features include at least: homogeneity, contrast, and entropy. Calculate the Euclidean distance from the target area cell to the nearest water source as a water potential correlation feature, and obtain the elevation, slope, and aspect of its location from digital elevation data as terrain features. Spectral features, spatial texture features, water potential correlation features, and terrain features are spliced ​​and standardized to construct a multi-source feature vector.

5. The method for rapid identification of vegetation species based on hyperspectral images according to claim 4, characterized in that, The construction of a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules specifically includes: Based on spectral feature parameters, spatial texture features, water potential correlation features, and terrain features, a node set of the knowledge graph is constructed. Based on statistical analysis and combined with expert experience rules, construct relationship edges between nodes; A knowledge graph for vegetation identification is constructed by using a graph database to store node sets and relational edges between nodes, where both nodes and relational edges contain quantitative attributes of their origin.

6. The method for rapid identification of vegetation species based on hyperspectral images according to claim 5, characterized in that, The step of constructing a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and outputting the probability of vegetation species in the target area and the overall classification confidence specifically includes: The adjacency matrix and node feature matrix are extracted from the vegetation recognition knowledge graph and used as the input matrix of the graph neural network. The adjacency matrix is ​​extracted from the relation edges and the node feature matrix is ​​extracted from the node set. Based on known vegetation species samples in the phenological observation network database, true category labels are set; Based on graph convolutional networks and temporal convolutional networks, a spatiotemporal graph convolutional network model is constructed. With the goal of predicting the deviation between the class label and the true class label, a loss function is constructed by weighted summation of cross-entropy loss and L2 regularization loss; Based on known vegetation species samples in the phenological observation network database, a knowledge graph of all vegetation species is drawn, and the corresponding adjacency matrix and node feature matrix are extracted. Based on the adjacency matrix and node feature matrix, the FGSM method is used to introduce training adversarial sample data to construct a vegetation recognition training sample set; Based on the vegetation recognition training sample set, a node feature attention mechanism and an adjacency attention mechanism are constructed, and multiple attention weights are formed by weighted summation to obtain the fused features. The fused features are input into a Softmax classification layer, which outputs the probability of vegetation species in the target area and the overall classification confidence.

7. The method for rapid identification of vegetation species based on hyperspectral images according to claim 6, characterized in that, The aforementioned multi-level vegetation identification mechanism, based on the spatial distribution of vegetation, specifically includes the identification and mapping of canopy, sub-canopy / understory vegetation, and ground cover vegetation, and includes: Satellites, drones, and land robots serve as sensing terminals, responsible for acquiring hyperspectral image data and transmitting the acquired data to the cloud platform; The cloud platform integrates a vegetation recognition knowledge graph and a spatiotemporal graph convolutional network model to form a vegetation recognition intelligent processing center. A multi-level vegetation identification mechanism is constructed, in which the cloud platform collaboratively processes data from different sensing terminals to identify and map vegetation at each level of the vertical structure.

8. A rapid vegetation species identification system based on hyperspectral images, characterized in that, A method for rapidly identifying vegetation species based on hyperspectral images as described in any one of claims 1-7 includes: The data acquisition module is used to acquire airborne and spaceborne hyperspectral image data of the target area, and simultaneously acquire digital elevation data, phenological information of the acquisition time, and climate zone information of the geographical location of the area. The data processing module is used to preprocess hyperspectral image data by radiometric calibration, atmospheric correction, and geometric correction to obtain a surface reflectance data cube; based on the surface reflectance data cube, it extracts the spectral features, spatial texture features, and correlation features with topography and water system of its internal pixels to construct a multi-source feature vector; based on the multi-source features, and combined with the phenological observation network database and expert experience rules, it constructs a vegetation recognition knowledge graph. The vegetation identification module is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. Based on the spatial distribution of vegetation, a multi-level vegetation identification mechanism is constructed to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively.

9. A rapid vegetation species identification system based on hyperspectral images according to claim 8, characterized in that, The data processing module includes: The data preprocessing unit is used to preprocess the hyperspectral image data by radiometric calibration, atmospheric correction and geometric correction to obtain a surface reflectance data cube; The feature extraction unit is used to extract the spectral features, spatial texture features, and correlation features with terrain and water system of the internal pixels of the surface reflectance data cube, and construct a multi-source feature vector. The knowledge graph unit is used to construct a vegetation identification knowledge graph based on multi-source features, a phenological observation network database, and expert experience rules.

10. A rapid vegetation species identification system based on hyperspectral images according to claim 9, characterized in that, The vegetation identification module includes: The vegetation identification unit is used to construct a spatiotemporal graph convolutional network model based on the vegetation identification knowledge graph, and output the probability of vegetation species in the target area and the overall classification confidence. A multi-level identification unit is used to construct a multi-level vegetation identification mechanism based on the spatial distribution of vegetation, and to identify and map the canopy, sub-canopy / understory vegetation and ground cover vegetation respectively.

Images