A method for identifying plants based on hyperspectral remote sensing of unmanned aerial vehicles

Abstract

The application discloses a kind of unmanned aerial vehicle-based hyperspectral remote sensing plant identification method, comprising: first, the route of unmanned aerial vehicle is passed through synchronous acquisition multi-source remote sensing data and is carried out high-precision registration, fusion three-dimensional information realizes the accurate segmentation of single plant, effectively solves the problem of crown adhesion and mixed pixel;On this basis, the pure spectrum and morphological characteristics of single plant are extracted, combined with hierarchical identification strategy and feature knowledge base, the fine identification from species to subspecies is completed;Through ground verification feedback, the results are corrected and the knowledge base is automatically updated, forming a closed loop of continuous optimization;The application significantly improves the recognition accuracy and automation level of plant monomer in complex vegetation community.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring technology, and in particular to a method for plant identification and recognition based on hyperspectral remote sensing using unmanned aerial vehicles (UAVs). Background Technology

[0002] Currently, using UAV remote sensing technology for plant identification has become an important tool in fields such as ecological surveys and agricultural monitoring. Existing technologies typically rely on hyperspectral or high-resolution visible light imagery, identifying vegetation communities through image classification or object-oriented methods. However, in densely vegetated areas, plant canopies often exhibit lateral contact and projection overlap in two-dimensional images, making it difficult to accurately segment individual plants. This mixed pixel problem not only renders the subsequently extracted canopy spectral features impure, severely affecting the accuracy of species identification, but also hinders morphological analysis of individual plants, fine subspecies identification, and quantitative research on ecological interactions between individuals. Furthermore, traditional methods often rely on a single data source or simple threshold segmentation, lacking effective utilization of the three-dimensional spatial structure of vegetation. They also exhibit insufficient generalization ability when facing complex community structures, and both automation and accuracy need improvement.

[0003] Existing UAV-based plant identification methods have significant limitations when facing complex real-world operating environments. On one hand, unavoidable platform vibrations and movements during UAV flight, along with plant swaying caused by natural wind, lead to motion blur and other image degradation in the acquired images, severely reducing image quality and significantly impacting the performance of segmentation and recognition algorithms based on the assumption of clear images. On the other hand, existing technical processes are mostly one-off identification surveys, ending once the results are output, lacking a closed-loop mechanism for continuous evolution of identification capabilities based on on-site verification feedback. This fails to fully utilize the spatial location and attribute information of identified individuals to further quantify and analyze the spatial competition relationships and growth dynamics within plant communities, limiting their in-depth application value in refined forest management, invasive species monitoring, and ecosystem process research. Summary of the Invention

[0004] This application provides a method for plant identification based on UAV hyperspectral remote sensing, which enables accurate segmentation of individual plants in complex vegetation communities. It effectively solves the problem of mixed pixels and canopy adhesion in traditional methods, and significantly improves the automation and identification accuracy of UAV remote sensing plant identification.

[0005] This application provides a method for plant identification and recognition based on hyperspectral remote sensing from unmanned aerial vehicles (UAVs), including: S1 plans the UAV flight path based on the terrain and vegetation density of the target area, and simultaneously / time-divisionally collects high-resolution visible light, hyperspectral and lidar data, and performs data preprocessing and geometric registration. S2 uses feature point matching and aerial triangulation technology to generate high-precision digital orthophotos and digital surface models from high-resolution visible light images. It performs geometric registration between hyperspectral data cubes and digital orthophotos, ensuring that each hyperspectral pixel corresponds one-to-one with the position on the visible light image. S3, based on a digital surface model, identifies potential single-plant canopy vertices as seed points by the local elevation change rate, generates superpixels by an improved simple linear iterative clustering algorithm that integrates color, texture and elevation information, and uses a graph optimization algorithm to segment single plants and output each single-plant mask. S4. Apply a single-plant segmentation mask to the hyperspectral image and obtain the average reflectance curve of all hyperspectral pixels within the mask. Extract various morphological and textural features of the plant from the high-resolution orthophoto image within the same mask area. S5. Construct a standard feature knowledge base containing species and subspecies that may appear in the target area. Species are determined by key band and spectral matching, and subspecies are identified by combining fine spectral and morphological features. S6 outputs a thematic map of single-plant identification results, guides ground verification of low-confidence targets, and uses the verification data to correct the results and update the knowledge base.

[0006] Preferably, the segmentation and output of each individual plant mask specifically includes: identifying potential individual plant canopy vertices as seed points based on a digital surface model; using the seed points as centers, performing an improved clustering algorithm that integrates color, texture, and elevation information on the orthophoto to generate superpixel blocks with homogeneous internal features; constructing a relational network graph by treating all superpixel blocks as nodes, where the weight of the connecting edges is determined by the color difference, texture difference, elevation difference, and spatial distance between adjacent superpixel blocks; applying a graph cut optimization algorithm to perform global optimization calculations on the network graph, merging superpixel blocks with similar features and separating superpixel blocks with dissimilar features to obtain preliminary individual plant regions; performing post-processing operations such as boundary smoothing and internal hole filling on the preliminary individual plant regions, and outputting a binarized mask file that precisely defines the pixel range of each independent plant.

[0007] Preferably, the subspecies identification combining fine spectral and morphological features includes: selecting fine feature bands for the species based on the species classification; calling the morphological and textural feature vectors of the individual plant; matching and comparing the fine spectral curve and morphological feature vector of the individual plant with reference templates of each subspecies under the species, wherein the matching process combines improved spectral information divergence and Euclidean distance of morphological features, and calculates a comprehensive similarity score according to a preset weight; judging the comprehensive similarity score according to a preset confidence threshold: if the score exceeds the high threshold, it is determined to be the subspecies; if the score is between the high and low thresholds, a list of several subspecies with the highest probability is output; if the scores are all below the low threshold, it is marked as an unknown subspecies or a suspected new variety.

[0008] Preferably, the method of determining species through key band and spectral matching includes: S51, constructing a Delaunay triangulation based on the center point coordinates of all individual plants to form a spatial adjacency graph; calculating the overlapping area of ​​adjacent canopies, determining the invasion relationship based on the overlap ratio threshold, and establishing directed competitive edges; S52, calculating the minimum convex hull and symmetry of the invaded canopy polygon, quantifying its morphological distortion degree through area ratio and symmetry index, and then calculating the competition pressure intensity index; S53, constructing a sub-network based on high-intensity competitive edges, identifying competitive clusters using a community detection algorithm, and ranking the dominance level within the cluster by calculating the net competition score to locate the regional range of the same species.

[0009] Preferably, the formation of the spatial adjacency graph specifically includes: generating a connection network based on the coordinates of the center points of all individual plants using the Delaunay triangulation algorithm, wherein each edge connects two spatially adjacent plant center points; and constructing an undirected adjacency graph with plants as nodes and connecting edges as adjacency relationships.

[0010] Preferably, the process of locating the region of the same species includes: constructing a strong competitive subnetwork based on the directed competitive edges and their corresponding competitive pressure intensities; running a community detection algorithm on the strong competitive subnetwork to identify a subset of tightly connected nodes as competitive clusters; calculating the net competitive score of each plant within each competitive cluster and identifying the dominance level based on the score ranking; and analyzing the species composition within each competitive cluster in conjunction with the species information of each plant, thereby locating the region where the same species is spatially concentrated and has significant internal competitive relationships.

[0011] Preferably, the method of locating the same species region further includes: S531, obtaining motion blur parameters that vary with depth based on high-frequency acquired flight state data and canopy height model, and constructing a spatially varying joint point spread function by fusing vibration model; S532, using a blind deconvolution algorithm that considers the spatial variability of the point spread function to restore the degraded image, and applying anisotropic filtering to suppress edge ringing artifacts; S533, constructing an end-to-end depth network that simultaneously estimates degradation parameters from state data, recovers features from degraded images, and directly outputs segmentation results based on the recovered features; performing high-precision registration on multi-frame overlapping degraded images, using a convex set projection algorithm for super-resolution reconstruction, and performing segmentation on the reconstructed high-resolution image; S534, calculating the comprehensive confidence score of segmentation and recognition based on the point spread function scale and texture degradation analysis of a single plant region, fusing the model's own probability entropy; estimating the image blur level in real time during flight, and dynamically adjusting the flight speed, altitude, and acquisition mode accordingly, while attempting to suppress identified vibration sources.

[0012] Preferably, obtaining the depth-varying motion blur parameters specifically involves: synchronously acquiring flight status data strictly aligned with the image exposure time during UAV flight; using a pre-obtained canopy height model to determine the canopy surface height corresponding to each pixel in the image; combining the camera imaging model and exposure time, and based on the flight status data and the height of each pixel, calculating the image-side motion vector of each pixel on the corresponding height plane caused by flight motion to generate a depth-varying spatial motion blur point spread function; performing spectral analysis on the acquired angular velocity data to identify and characterize the periodic vibration of the aircraft itself and the random jitter components caused by environmental disturbances, generating corresponding vibration point spread functions; and fusing the spatially varied motion blur point spread function with the vibration point spread function to form a spatially varied joint point spread function describing the degradation process of each pixel.

[0013] Preferably, the segmentation on the reconstructed high-resolution image includes: applying a graph cut algorithm to segment the image based on the reconstructed high-resolution image, wherein the energy function of the graph cut algorithm includes a data term and a smoothing term; the data term utilizes the spectral and gradient information of the image, and the smoothing term is used to encourage spatially adjacent and feature-similar pixels to belong to the same label; detecting the boundaries obtained after segmentation and analyzing their texture sources to identify and correct artifact boundaries generated by the super-resolution reconstruction process.

[0014] Preferably, the high-precision registration of the multi-frame overlapping degraded images specifically includes: performing radiometric consistency correction and denoising preprocessing on the multi-frame overlapping image sequence; estimating the inter-frame motion field using an improved variational optical flow method that introduces the gradient constant assumption and semantic segmentation anisotropic smoothing weights; and registering the multi-frame images to a unified reference coordinate system based on the inter-frame motion field.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: By fusing multi-source data and using 3D information, precise segmentation of individual plants in complex vegetation communities was achieved, effectively solving the problem of mixed pixels and canopy adhesion in traditional methods, thus extracting pure individual spectra. Based on this, a hierarchical identification strategy was adopted, which not only accurately identifies species but also finely distinguishes subspecies. Guided ground verification and self-updating of the knowledge base provide continuous learning and increasingly accurate capabilities, significantly improving the automation and identification accuracy of UAV remote sensing plant identification.

[0016] Building upon the ability to accurately identify individual plants, this technology further overcomes the limitations of traditional remote sensing, which can only identify plant types. It achieves dynamic quantification and visual analysis of spatial competition relationships among plants, automatically identifying competitive hotspots and key dominant / vulnerable individuals within the community. This technology transforms ecological interactions that are difficult to observe directly into objective and measurable spatial competition indices, providing direct and actionable scientific decision-making support for precision forest tending, invasive plant control, and the protection of rare species. Furthermore, it can predict community succession trends based on the current competitive landscape, achieving a leap from static surveys to dynamic relationship management and prediction.

[0017] By jointly estimating degradation parameters and restored image features using deep networks, and combining this with multi-frame super-resolution reconstruction, the segmentation accuracy and robustness of single plants under adverse conditions such as blurring and aliasing are significantly improved. Real-time image quality evaluation and dynamic optimization of the flight acquisition strategy ultimately output high-precision segmentation and recognition results with credibility assessment, providing a more reliable data foundation for subsequent ecological analysis.

[0018] By accurately decoupling plant self-motion from a mixed motion field using Kalman filtering and establishing a physical driving model, degradation caused by both UAV motion and plant self-motion is eliminated during image reconstruction. This significantly improves the accuracy and reliability of single-plant segmentation in windy conditions and contributes a novel feature dimension of motion phenotype to species identification, ultimately greatly enhancing robustness and information acquisition capabilities under complex natural conditions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for identifying and classifying plants using hyperspectral remote sensing based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Detailed Implementation

[0020] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terminology used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1: Figure 1 This is a flowchart illustrating a method for identifying and classifying plants using hyperspectral remote sensing based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention.

[0023] like Figure 1 As shown, a method for plant identification based on UAV hyperspectral remote sensing includes the following steps: S1 plans the drone flight path based on the terrain and vegetation density of the target area, and simultaneously / time-divisionally collects high-resolution visible light, hyperspectral and lidar data, and performs data preprocessing and geometric registration.

[0024] Among them, high-resolution oblique photography acquires visible light multi-angle images with a spatial resolution better than 2 cm, which are used for subsequent 3D information extraction and fine segmentation; hyperspectral imaging acquires hyperspectral data cubes with a spectral range of 400-2500 nanometers and a spectral resolution better than 10 nanometers, which are used for subspecies-level spectral feature analysis; and lidar acquires high-precision point cloud data, which are used to assist in verifying segmentation boundaries and extracting canopy height models.

[0025] S2 employs feature point matching and aerial triangulation techniques to generate high-precision digital orthophotos and digital surface models from high-resolution visible light images. It performs geometrical registration between the hyperspectral data cube and the digital orthophoto, ensuring that each hyperspectral pixel corresponds one-to-one with the position on the visible light image.

[0026] S3, based on a digital surface model, identifies potential single-plant canopy vertices as seed points by using the local elevation change rate. It generates superpixels by using an improved simple linear iterative clustering algorithm that integrates color, texture, and elevation information, and then uses a graph optimization algorithm to segment single plants and output a mask for each single plant.

[0027] Specifically, based on a digital surface model, the system automatically identifies prominent, independent high points on the ground surface by analyzing the elevation variations of each local area. These high points are identified as potential tops of individual plant canopies and used as the starting center points for segmentation.

[0028] Using the aforementioned seed points as centers, an improved clustering segmentation is performed on high-resolution orthophotos. When classifying adjacent pixels, this algorithm considers not only color and texture similarity but also their elevation proximity. In this way, a large number of small, homogeneous patches with consistent spectral, textural, and spatial elevation characteristics are generated, laying the foundation for distinguishing different plants.

[0029] A relational network graph is constructed by treating all small blocks as basic units. The strength of the connecting lines in the graph is determined by the differences in color, texture, height, and actual distance between adjacent blocks. A graph cut optimization algorithm is used to perform global optimization calculations on this network based on the principle that blocks with similar features should be merged and blocks with dissimilar features should be separated. All blocks belonging to the same plant are merged and completely separated from blocks of adjacent plants, thus obtaining the initial single-plant regions.

[0030] Post-processing operations such as contour smoothing and internal hole repair are performed on each plant region extracted in the above steps to make its boundaries more precise and the region more complete. A corresponding closed binary mask file is output for each individual plant, which accurately marks all pixel positions of the plant in the image.

[0031] S4. Apply a single-plant segmentation mask to the hyperspectral image and obtain the average reflectance curve of all hyperspectral pixels within the mask. Within the same mask area, extract various morphological and textural features of the plant from the high-resolution orthophoto image.

[0032] Specifically, the binarized mask of the single plant generated in the previous step is strictly mapped onto the hyperspectral image that has undergone geometric registration. All hyperspectral pixels falling within the mask's range are automatically identified and selected. For all selected hyperspectral pixels, the average reflectance is calculated for each band. This series of averages is connected in wavelength order to form a pure average reflectance curve representing the overall spectral characteristics of the single plant's canopy.

[0033] Based on a single plant mask, the geometric properties of the region defined by the mask are automatically calculated on a high-resolution digital orthophoto. This includes the projected area, perimeter, and the lengths of the major and minor axes of the minimum bounding rectangle. This leads to the derivation of various parameters describing its macroscopic morphology, such as aspect ratio, shape compactness, and ellipticity. Texture analysis is then performed on the corresponding region of the orthophoto (usually converted to grayscale) using methods such as gray-level co-occurrence matrix. By setting distance and orientation parameters, statistical quantities such as contrast, homogeneity, energy, and entropy of the region are calculated, thereby quantifying the texture roughness, uniformity, and regularity of the canopy surface.

[0034] S5. Construct a standard feature knowledge base containing species and subspecies that may appear in the target area. Species are identified by key band and spectral matching, and subspecies are identified by combining fine spectral and morphological features.

[0035] Specifically, a structured reference feature knowledge base is established. This knowledge base systematically collects standard data on species and their subspecies that may appear within the target area. Each record includes: subspecies identification, average spectral curves of the pure canopy collected at different typical phenological stages (such as leaf unfolding, flowering, and fruiting), statistical ranges of typical morphological features (such as area and aspect ratio), ranges of textural features, and corresponding habitat information. The knowledge base data comes from historical ground measurements or authoritatively confirmed prior UAV remote sensing samples.

[0036] For the complete spectral dataset of all species in the knowledge base, methods such as the continuous projection algorithm were used for analysis to select a set of key spectral band combinations that are most sensitive and discriminative in distinguishing different species. The response values ​​of the spectral curves of the individual plants to be identified on the key band subset were extracted and spectral angle matching calculations were performed with the reference spectral templates of each species in the knowledge base. The species with the smallest spectral angle, i.e., the most similar spectral shape, was preliminarily determined as the species affiliation of the individual plant.

[0037] Intra-species fine feature selection: Within the identified species range, a pre-defined secondary identification model for that species is activated. This model uses another set of optimized and selected, more refined feature bands, which typically sensitively reflect differences in the biochemical components within leaves to distinguish subspecies. Simultaneously, the morphological and textural feature vectors extracted from the individual plant are utilized. The fine spectral curve and morphological features of the plant to be identified are compared with templates from all subspecies within that species. The matching process combines improved spectral information divergence (measuring spectral similarity) and Euclidean distance of morphological features, and performs a comprehensive score according to pre-defined weights.

[0038] Preset high and low confidence thresholds. If the overall similarity score of the best matching subspecies exceeds the high threshold, it is directly identified as that subspecies; if the score is between the high and low thresholds, a list of several subspecies with the highest probability is output; if the scores are all below the low threshold, it is marked as an unknown subspecies or a suspected new variant, and its characteristics are recorded for subsequent research.

[0039] S6 outputs a thematic map of single-plant identification results, guides ground verification of low-confidence targets, and uses the verification data to correct the results and update the knowledge base.

[0040] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By fusing multi-source data and using 3D information, precise segmentation of individual plants in complex vegetation communities was achieved, effectively solving the problem of mixed pixels and canopy adhesion in traditional methods, thus extracting pure individual spectra. Based on this, a hierarchical identification strategy was adopted, which not only accurately identifies species but also finely distinguishes subspecies. Guided ground verification and self-updating of the knowledge base provide continuous learning and increasingly accurate capabilities, significantly improving the automation and identification accuracy of UAV remote sensing plant identification.

[0041] Example 2: In Example 1, the fusion of high-resolution imagery and hyperspectral data enabled precise segmentation of individual plants and detailed identification at the subspecies level, and constructed a knowledge base containing spectral and morphological characteristics of species / subspecies. However, the output of this method mainly presents a static survey list and distribution map, failing to further reveal the spatial interactions between individual plants and the resulting differences in growth status. In practical applications such as forest management and ecological protection, managers need to know which plants are being strongly suppressed, where the competition hotspots are, and which key individuals can be removed to achieve the greatest ecological benefits with minimal intervention. To move from static identification to dynamic relationship analysis and growth intervention simulation, it is necessary to quantify and visualize the spatial competition relationships between plants. Therefore, this example makes the following optimizations and improvements based on the precise individual plant information provided in Example 1.

[0042] In some embodiments, species are determined by matching key bands with spectra, and step S5 further includes: S51. Based on the coordinates of the center point of all individual plants, construct the Delaunay triangulation to form a spatial adjacency graph; calculate the overlapping area of ​​adjacent canopies, determine the invasion relationship based on the overlap ratio threshold, and establish directed competing edges.

[0043] Specifically, based on the geographic coordinates of the center points of all individual plants, the Delaunay triangulation algorithm is applied to generate a connected network composed of triangles. Each edge of this network connects two spatially adjacent plant center points, forming an undirected adjacency graph with plants as nodes and spatial adjacency relationships as connecting edges.

[0044] Detecting Canopy Invasion and Determining Competition Direction: For each edge in the adjacency graph (i.e., each pair of adjacent plants, denoted as plant A and plant B): a. Calculate the canopy overlap area: Calculate the area of ​​the overlapping region between the precise two-dimensional canopy polygons of plant A and plant B using polygon Boolean operations. b. Determine the existence of competition: If the overlap area is simultaneously greater than a preset threshold (e.g., 5%) for the canopy areas of both plant A and plant B, a significant competition relationship is determined between them. Otherwise, this edge is ignored. c. Determine the competition direction: For competing plant pairs, calculate the proportion of the overlapping area belonging to plant A and plant B, respectively. If the proportion of the area from plant A is significantly higher than that from plant B (e.g., >60%), then plant A is defined as the invading plant and plant B as the invaded plant, and a directed edge is established from the invading plant to the invaded plant. Finally, an initial directed network of plant competition relationships is determined. In the graph, nodes represent individual plants, and directed edges represent significant unidirectional spatial competition relationships from intruders to intruded plants.

[0045] S52 calculates the minimum convex hull and symmetry of the intruded canopy polygon, quantifies its morphological distortion degree by area ratio and symmetry index, and then calculates the competitive compression intensity index.

[0046] Specifically, for each canopy polygon of an invaded plant, its minimum convex hull (i.e., the convex polygon with the smallest area that completely encloses all vertices of the polygon) is calculated, and this is approximated as the ideal canopy shape of the plant under no competitive pressure. Canopy integrity is calculated using the formula: Canopy integrity = Actual canopy polygon area / Minimum convex hull area. The smaller this value is than 1, the greater the area loss due to competition.

[0047] Calculate the actual canopy polygon area and the area of ​​the smallest convex hull of the invaded plant, and then calculate the canopy integrity index using the formula: Canopy Integrity = Actual Canopy Area / Convex Hull Area. This index reflects the loss of canopy projected area due to competitive pressure. The closer the value is to 1, the more intact the morphology; the smaller the value (e.g., 0.7), the more severe the morphological distortion, indicating that approximately 30% of the potential growth space has been occupied by adjacent plants.

[0048] Divide the polygon into m directions (e.g., m=12) with the centroid of the canopy polygon as the center, and calculate the radial distance from the centroid to the polygon boundary in each direction. The symmetry index S is calculated using the following formula:

[0049] Where n = m / 2, and These represent the radial distances in the k-th direction and its opposite direction (180 degrees apart). The closer the S value is to 1, the more symmetrical the shape; the closer it is to 0, the stronger the asymmetry caused by unilateral compression.

[0050] Calculate the competitive pressure intensity index for each pair of established competitive relationships (intruder A and intruded party B). Calculate and normalize the following four core factors within the interval [0,1]: Relative size advantage: The ratio of the canopy area of ​​the invading party A to the canopy area of ​​the invaded party B. Intrusion degree: The ratio of the overlapping area of ​​the canopies of both parties to the canopy area of ​​the intruded party B. Morphological distortion degree: 1 minus the canopy integrity of the invaded side B. Spatial proximity: The ratio of the distance between the center points of two plants to the sum of their canopy radii (estimated by equivalent radius). The equivalent radius of a single plant's canopy can be calculated from its canopy area, i.e., R = sqrt(area / π).

[0051] The normalized factor values ​​are then assigned preset weights. Perform a weighted summation: This yields a comprehensive index CPI ranging from 0 to 1, which is used to quantitatively characterize the intensity of unidirectional competitive pressure from A to B.

[0052] S53 constructs a sub-network based on high-intensity competitive edges, uses a community detection algorithm to identify competitive clusters, and ranks the dominance level within clusters by calculating net competition scores to locate the regional range of the same species.

[0053] Specifically, in the directed competitive network generated by S51, only edges with a competitive pressure intensity index (CPI) exceeding a preset intensity threshold are retained to construct a strongly competitive subnetwork. A community detection algorithm is then applied to this strongly competitive subnetwork to identify subsets of nodes that are tightly connected but relatively sparsely connected to external nodes as competitive clusters. Each competitive cluster represents a locally highly competitive spatial region.

[0054] Within each competitive cluster, for each plant, its net competition score (number of times as an intruder minus number of times as a victim) is calculated. All plants are ranked according to their scores, with the plant with the highest score identified as the dominant individual (primary source of oppression) within that cluster, and the plant with the lowest score or a negative score identified as the oppressed individual. Combining this with the species information of each plant identified in Example 1, the species composition within each competitive cluster can be further analyzed, and specific areas where the same species is spatially concentrated and internally highly competitive can be located.

[0055] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: Building upon the ability to accurately identify individual plants, this technology further overcomes the limitations of traditional remote sensing, which can only identify plant types. It achieves dynamic quantification and visual analysis of spatial competition relationships among plants, automatically identifying competitive hotspots and key dominant / vulnerable individuals within the community. This technology transforms ecological interactions that are difficult to observe directly into objective and measurable spatial competition indices, providing direct and actionable scientific decision-making support for precision forest tending, invasive plant control, and the protection of rare species. Furthermore, it can predict community succession trends based on the current competitive landscape, achieving a leap from static surveys to dynamic relationship management and prediction.

[0056] Example 3: In Example 2, the spatial adjacency and canopy invasion relationships among identified individual plants were analyzed to achieve a quantitative assessment and visualization of the plant competition pattern. However, the reliability of this analysis result is entirely based on the high-precision individual plant segmentation mask provided in Example 1. When the UAV flies in complex operating environments (such as high wind disturbance and high-density vegetation areas), image degradation caused by platform vibration and motion (such as motion blur and vibration shadows) will significantly reduce the core segmentation accuracy of Example 1, thus rendering the competition relationship analysis in Example 2 untenable. Faced with image degradation of different degrees and types, if only a unified segmentation model designed for ideal imaging conditions is used, its generalization ability will inevitably be insufficient, and the segmentation error will directly transmit and distort the subsequent ecological relationship analysis results. In order to ensure that reliable individual plant segmentation and attribute information can still be obtained under complex imaging conditions, thereby providing a solid data foundation for the competition analysis in Example 2 and even the identification results in Example 1, it is necessary to perform collaborative optimization and adaptive enhancement of the degraded image restoration process, the segmentation model itself, and the flight acquisition strategy.

[0057] In some embodiments, locating the region of the same species in step S53 further includes: S531, based on high-frequency acquired flight state data and canopy height model, obtains motion fuzzy parameters that vary with depth, and integrates vibration model to construct spatially varying joint point spread function.

[0058] Specifically, during the drone's flight, the inertial measurement unit's three-axis acceleration and three-axis angular velocity, the fused flight speed, and the rotational speed of each motor are synchronously acquired at a frequency of no less than 200Hz, with all data strictly aligned with the image exposure time. Subsequently, using a pre-obtained canopy height model, each pixel in the image is assigned its corresponding true canopy surface height value. Combining the camera imaging model and exposure time, the image-side motion vector caused by the flight speed on the corresponding height plane is independently calculated for each pixel, thus obtaining a spatially variable motion blur point spread function that varies with target depth. Simultaneously, spectral analysis is performed on the gyroscope angular velocity data to identify the periodic vibration components caused by the motor rotation fundamental frequency and its harmonics, as well as broadband random jitter components caused by wind disturbance, etc., and these are mathematically represented using sinusoidal modulation models and Gaussian models, respectively. Finally, the calculated spatially variable motion blur point spread function is convolved and fused with the point spread function representing vibration artifacts to form a spatially variable joint point spread function describing the final degradation process of each pixel.

[0059] S532 employs a blind deconvolution algorithm that considers the spatial variability of the point spread function to restore degraded images, and applies anisotropic filtering to suppress edge ringing artifacts.

[0060] S533 constructs an end-to-end deep network that simultaneously estimates degradation parameters from state data, recovers features from degraded images, and directly outputs segmentation results based on the recovered features; performs high-precision registration on multi-frame overlapping degraded images, uses the convex set projection algorithm for super-resolution reconstruction, and performs segmentation on the reconstructed high-resolution image.

[0061] Specifically, an end-to-end deep neural network is constructed, with inputs consisting of a single-frame degraded image and synchronously acquired flight state time-series data. The network comprises three tightly coupled sub-modules: a degradation parameter estimation sub-module, which takes flight state data (acceleration, angular velocity, etc.) as input and, through an encoder network, predicts a compact set of parameters used to parameterize the joint point spread function. A feature recovery sub-module, taking the degraded image and the estimated point spread function parameters as input, contains a series of deformable convolutional or adaptive kernel prediction layers. These layers utilize the predicted point spread function parameters to guide the generation of convolutional kernels, thereby performing targeted deblurring and de-shading operations at the feature level and outputting a restored, sharp feature map. A segmentation sub-module, taking the restored sharp feature map as input, directly outputs the class probability of each pixel—a preliminary plant segmentation mask—through an encoder-decoder structured semantic segmentation network (such as a variant of U-Net).

[0062] The network employs a joint loss function for end-to-end training. This function is a weighted sum of segmentation losses (such as cross-entropy loss or Dice loss) and an auxiliary restoration perception loss (such as feature matching loss). Training data is obtained by simulating a degradation data generation process. This involves using clear hyperspectral-visible images and their ground truth segmentation values, combined with the parameter distribution from real flight data statistics, to synthesize a large number of training samples with physically realistic degradation.

[0063] For multiple frames of degraded images with sufficient overlap acquired from the same scene, scale-invariant feature transformation (SIN) points are extracted from each frame for matching and mismatching. By solving the homography transformation matrix, all frames are precisely registered to the same reference coordinate system, achieving sub-pixel alignment. Based on the registration, a convex set projection algorithm is employed. The initial high-resolution grid estimate (obtained by bicubic interpolation) is iteratively projected onto multiple convex constraint sets, including observation consistency constraints (the reconstructed image should approximate the original low-resolution observation after degradation and downsampling), energy bounded constraints, non-negativity constraints, and edge sparsity constraints. Iteration continues until convergence, resulting in a high-resolution image with enhanced details. On the reconstructed high-resolution image, a graph cut algorithm is applied for fine segmentation. The energy function's data term utilizes image spectral and gradient information, while the smoothing term encourages spatially adjacent and feature-similar pixels to belong to the same label. After segmentation, a crucial post-processing step is performed: detecting segmentation boundaries and analyzing their texture origins. If a certain boundary segment mainly originates from false high-frequency textures that do not exist in any of the original low-resolution images but are generated only during the reconstruction process, it is determined to be a super-resolution artifact. The boundary segment is then smoothed or corrected based on information from adjacent frames to ensure that the segmentation results are based on real observation information.

[0064] S534, based on the point spread function scale and texture degradation analysis of a single plant region, integrates the model's own probability entropy to calculate the comprehensive confidence score for segmentation and recognition; it estimates the image blur level in real time during flight and dynamically adjusts the flight speed, altitude and acquisition mode accordingly, while attempting to suppress the identified vibration sources.

[0065] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By jointly estimating degradation parameters and restored image features using deep networks, and combining this with multi-frame super-resolution reconstruction, the segmentation accuracy and robustness of single plants under adverse conditions such as blurring and aliasing are significantly improved. Real-time image quality evaluation and dynamic optimization of the flight acquisition strategy ultimately output high-precision segmentation and recognition results with credibility assessment, providing a more reliable data foundation for subsequent ecological analysis.

[0066] Example 4: Example 3 primarily addressed the degradation caused by the motion of the UAV platform. However, under windy conditions, the movement of the plants themselves can also cause severe image blurring, and this blurring is spatially intertwined with the platform motion blur, making it impossible to separate using traditional methods. This example, through high-frame-rate synchronous observation, robust optical flow estimation, and physical-driven motion decoupling and modeling, not only can a clearer image be recovered from the mixed blur, but it can also extract the unique motion phenotypic features of the plants. This improves segmentation accuracy while providing a new discriminative dimension for species identification.

[0067] In some embodiments, step S533, which involves high-precision registration of multiple overlapping degraded images, further includes: 3A, through the synchronous trigger module, adds a high-precision timestamp to the data streams acquired by all sensors, and resamples and aligns the inertial measurement unit data to the center exposure time of each frame image.

[0068] 3B. After radiometric correction and denoising of high-speed image sequences, an improved variational optical flow method is used for motion estimation. This method introduces the gradient constant assumption and anisotropic smoothing weights based on semantic segmentation to adapt to the appearance changes and motion discontinuities of vegetation scenes.

[0069] 3C, based on accurate sensor calibration parameters and canopy height model, establishes a projection chain from the global coordinate system to the image plane, and calculates the image plane velocity vector field caused by UAV motion that varies with canopy height.

[0070] In 3D, a Kalman filter is established for each pixel. The drone's motion speed calculated by projection is used as the control input, and the optical flow observation value is used as the measurement to decouple the plant's own motion speed. Principal component analysis is used to separate its overall swaying and leaf trembling components.

[0071] 3E: Based on the extracted swing frequency and amplitude parameters of a single plant, its simple harmonic motion trajectory is simulated to generate a point spread function describing the fuzzy reciprocating motion. This function is then convolved with a Gaussian point spread function characterizing leaf tremor to construct a plant self-movement degradation model.

[0072] 3F, by convolving the drone motion point spread function with the plant self-motion point spread function, constructs a joint degradation model of spatial variation, and by solving a large-scale linear inverse problem, deconvolution is performed to obtain a restored image that simultaneously eliminates the two types of motion blur.

[0073] For example, in the case study of windy observation of a poplar forest, assuming a drone is used to observe a mixed poplar and birch forest under wind speed conditions of 3-4.

[0074] The drone captured images at 120 frames per second, and the inertial measurement unit recorded them at 200 Hz. At the center exposure time t=10.253 seconds of a certain frame, the system interpolated to find the precise drone attitude and velocity at that moment, which served as the benchmark for motion analysis of that frame.

[0075] In the image sequence, poplar leaves flip (abrupt brightness changes), but the improved optical flow method, through the assumption of constant gradient, can still accurately track their motion. Meanwhile, the crowns of adjacent poplar and birch trees sway in slightly different directions in the wind. The anisotropic smoothing term based on semantic segmentation allows for the existence of this motion discontinuity, thus obtaining an accurate mixed motion field.

[0076] It is known that the height of the poplar tree at the top of the canopy is approximately 15 meters, and the height of the birch tree at the bottom is approximately 8 meters. Based on the attitude and velocity of the UAV at t=10.253 seconds, the image displacement vector of the pixel at the 15-meter height plane due to the UAV's forward movement is calculated to be (2.5,0) pixels, while the image displacement vector at the 8-meter height plane is (4.7,0) pixels, confirming the depth dependence of motion blur.

[0077] For a single pixel in the canopy of a poplar tree, the observed total optical flow is (15.2, 3.1). Step 3C calculates the UAV motion contribution at this point as (2.5, 0). After Kalman filtering decoupling, the estimated plant motion at this pixel is (12.7, 3.1). Analyzing the motion of all pixels of this poplar tree reveals that it mainly oscillates at a frequency of approximately 0.8 Hz in the east-west direction with an amplitude of approximately ±5 pixels, while simultaneously superimposed with high-frequency, small-amplitude random leaf tremors.

[0078] For this poplar tree, its east-west reciprocating motion trajectory within a single frame's 20-millisecond exposure time was simulated using its 0.8 Hz, ±5 pixel sway parameters. The generated point spread function exhibits a unique bimodal morphology. Convolving this with a Gaussian kernel simulating leaf tremor yields the tree's final self-motion degradation kernel.

[0079] The self-motion degradation kernel of this poplar tree is convolved with the drone motion degradation kernel calculated in step 3C to obtain its composite point spread function. Solving this composite point spread function on a large linear system consisting of all trees in the entire image (each with its own unique composite point spread function) yields a clear image that simultaneously eliminates both drone trailing and tree swaying ghosting. From the poplar tree's motion, its dominant swaying frequency of 0.8 Hz can also be extracted as a feature, while the adjacent birch tree's swaying frequency is 1.2 Hz, providing new clues for species identification.

[0080] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: By accurately decoupling plant self-motion from a mixed motion field using Kalman filtering and establishing a physical driving model, degradation caused by both UAV motion and plant self-motion is eliminated during image reconstruction. This significantly improves the accuracy and reliability of single-plant segmentation in windy conditions and contributes a novel feature dimension of motion phenotype to species identification, ultimately greatly enhancing robustness and information acquisition capabilities under complex natural conditions.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for plant identification and recognition based on hyperspectral remote sensing using unmanned aerial vehicles (UAVs), characterized in that, include: S1 plans the UAV flight path based on the terrain and vegetation density of the target area, and simultaneously / time-divisionally collects high-resolution visible light, hyperspectral and lidar data, and performs data preprocessing and geometric registration. S2 uses feature point matching and aerial triangulation technology to generate high-precision digital orthophotos and digital surface models from high-resolution visible light images. It performs geometric registration between hyperspectral data cubes and digital orthophotos, ensuring that each hyperspectral pixel corresponds one-to-one with the position on the visible light image. S3, based on a digital surface model, identifies potential single-plant canopy vertices as seed points by the local elevation change rate, generates superpixels by an improved simple linear iterative clustering algorithm that integrates color, texture and elevation information, and uses a graph optimization algorithm to segment single plants and output each single-plant mask. S4. Apply a single-plant segmentation mask to the hyperspectral image and obtain the average reflectance curve of all hyperspectral pixels within the mask. Extract various morphological and textural features of the plant from the high-resolution orthophoto image within the same mask area. S5. Construct a standard feature knowledge base containing species and subspecies that may appear in the target area. Species are determined by key band and spectral matching, and subspecies are identified by combining fine spectral and morphological features. S6 outputs a thematic map of single-plant identification results, guides ground verification of low-confidence targets, and uses the verification data to correct the results and update the knowledge base.

2. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 1, characterized in that, The segmentation and output of each individual plant mask specifically includes: identifying potential individual plant canopy vertices as seed points based on a digital surface model; using the seed points as centers, performing an improved clustering algorithm that integrates color, texture, and elevation information on the orthophoto to generate superpixel blocks with homogeneous internal features; constructing a relational network graph by treating all superpixel blocks as nodes, where the weight of the connecting edges is determined by the color difference, texture difference, elevation difference, and spatial distance between adjacent superpixel blocks; applying a graph cut optimization algorithm to perform global optimization calculations on the network graph, merging superpixel blocks with similar features and separating superpixel blocks with dissimilar features to obtain preliminary individual plant regions; and performing post-processing operations such as boundary smoothing and internal hole filling on the preliminary individual plant regions to output a binarized mask file that precisely defines the pixel range of each independent plant.

3. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 1, characterized in that, The subspecies identification method combining fine spectral and morphological features includes: selecting fine feature bands for the species based on its classification; calling the morphological and textural feature vectors of the individual plant; matching and comparing the fine spectral curve and morphological feature vector of the individual plant with reference templates for each subspecies under the species, wherein the matching process combines improved spectral information divergence and Euclidean distance of morphological features, and calculates a comprehensive similarity score according to preset weights; judging the comprehensive similarity score according to a preset confidence threshold: if the score exceeds the high threshold, it is determined to be the subspecies; if the score is between the high and low thresholds, a list of several subspecies with the highest probability is output; if the scores are all below the low threshold, it is marked as an unknown subspecies or a suspected new variety.

4. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 1, characterized in that, The method of identifying species through key band and spectral matching includes: S51, constructing a Delaunay triangulation based on the center point coordinates of all individual plants to form a spatial adjacency graph; calculating the overlapping area of ​​adjacent canopies, determining the invasion relationship based on the overlap ratio threshold, and establishing directed competitive edges; S52, calculating the minimum convex hull and symmetry of the invaded canopy polygon, quantifying its morphological distortion degree through area ratio and symmetry index, and then calculating the competition pressure intensity index; S53, constructing a sub-network based on high-intensity competitive edges, identifying competitive clusters using a community detection algorithm, and ranking the dominance level within the cluster by calculating the net competition score to locate the regional range of the same species.

5. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 4, characterized in that, The formation of the spatial adjacency graph specifically includes: generating a connection network based on the coordinates of the center points of all individual plants using the Delaunay triangulation algorithm, wherein each edge connects two spatially adjacent plant center points; and constructing an undirected adjacency graph with plants as nodes and connecting edges as adjacency relationships.

6. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 4, characterized in that, The process of locating the region of the same species includes: constructing a strong competitive subnetwork based on the directed competitive edges and their corresponding competitive pressure intensities; running a community detection algorithm on the strong competitive subnetwork to identify a subset of tightly connected nodes as competitive clusters; calculating the net competitive score of each plant within each competitive cluster and identifying the dominance level based on the score ranking; and analyzing the species composition within each competitive cluster in conjunction with the species information of each plant to locate the region where the same species is spatially concentrated and has significant internal competitive relationships.

7. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 4, characterized in that, The method for locating the same species region also includes: S531, obtaining motion blur parameters that vary with depth based on high-frequency acquired flight state data and canopy height model, and constructing a spatially varying joint point spread function by fusing vibration model; S532, using a blind deconvolution algorithm that considers the spatial variability of the point spread function to restore the degraded image, and applying anisotropic filtering to suppress edge ringing artifacts; S533, constructing an end-to-end depth network that simultaneously estimates degradation parameters from state data, recovers features from degraded images, and directly outputs segmentation results based on the recovered features; performing high-precision registration on multi-frame overlapping degraded images, using a convex set projection algorithm for super-resolution reconstruction, and performing segmentation on the reconstructed high-resolution image; S534, calculating the comprehensive confidence score of segmentation and recognition based on the point spread function scale and texture degradation analysis of a single plant region, fusing the model's own probability entropy; estimating the image blur level in real time during flight, and dynamically adjusting flight speed, altitude, and acquisition mode accordingly, while attempting to suppress identified vibration sources.

8. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 7, characterized in that, The motion blur parameters that vary with depth are obtained as follows: During the flight of the UAV, flight status data that is strictly aligned with the image exposure time is collected synchronously; using a pre-obtained canopy height model, the canopy surface height corresponding to each pixel in the image is determined; combining the camera imaging model and the exposure time, based on the flight status data and the height of each pixel, the image-side motion vector of each pixel on the corresponding height plane caused by flight motion is calculated to generate a spatially variable motion blur point spread function that varies with depth; spectral analysis is performed on the collected angular velocity data to identify and characterize the periodic vibration of the aircraft itself and the random jitter components caused by environmental disturbances, and corresponding vibration point spread functions are generated. The spatially variable motion blur point spread function is fused with the vibration point spread function to form a spatially variable joint point spread function that describes the degradation process of each pixel.

9. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 7, characterized in that, The segmentation process on the reconstructed high-resolution image includes: applying a graph cut algorithm to segment the image based on the reconstructed high-resolution image, wherein the energy function of the graph cut algorithm includes a data term and a smoothing term; the data term utilizes the spectral and gradient information of the image, and the smoothing term is used to encourage spatially adjacent and feature-similar pixels to belong to the same label; detecting the boundaries obtained after segmentation and analyzing their texture sources to identify and correct artifact boundaries generated by the super-resolution reconstruction process.

10. The method for plant identification and recognition based on UAV hyperspectral remote sensing according to claim 7, characterized in that, The high-precision registration of multi-frame overlapping degraded images specifically includes: performing radiometric consistency correction and denoising preprocessing on the multi-frame overlapping image sequence; estimating the inter-frame motion field using an improved variational optical flow method that introduces the gradient constant assumption and semantic segmentation anisotropic smoothing weights; and registering the multi-frame images to a unified reference coordinate system based on the inter-frame motion field.

Images