A method and system for monitoring the status of garden greening based on UAV remote sensing

Abstract

This application discloses a method and system for monitoring the status of garden greening based on UAV remote sensing, belonging to the field of image processing technology. The method includes: acquiring point cloud and hyperspectral images of the garden; generating a preliminary high-density LiDAR point cloud; identifying absolute blind spots, performing pixel decomposition on pixels in the absolute blind spots, inferring the morphological parameters of vegetation in the blind spots, and stitching the point clouds to obtain a complete high-density LiDAR point cloud; generating SfM sparse point clouds and fusing them to form a fused point cloud; extracting multi-dimensional features, inputting them into an improved ConvLSTM model, and obtaining a full-area vegetation feature set including growth and aesthetics; and performing growth assessment and aesthetics assessment. This application solves the blind spot problem caused by canopy occlusion by using dual-band LiDAR multi-view fusion, hyperspectral hybrid pixel decomposition, and point cloud completion technology; at the same time, it constructs a dual-dimensional assessment system of growth and aesthetics to avoid the problem of disconnect from the ground experience caused by focusing only on growth.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method for monitoring the status of vegetation at different levels in a garden using laser point cloud and hyperspectral images. Background Technology

[0002] With the acceleration of urbanization, urban greening, as a core component of the urban ecosystem, is experiencing continuous growth in construction scale and investment, placing higher demands on the efficiency, accuracy, and comprehensiveness of monitoring technologies. Traditional urban greening monitoring relies on manual field surveys, which suffers from inherent drawbacks such as low efficiency, high cost, strong subjectivity, and poor timeliness, making it difficult to meet the needs of large-area, high-frequency monitoring. Unmanned aerial vehicle (UAV) remote sensing technology, with its advantages of flexibility, wide coverage, and rapid data acquisition, has become the mainstream technology for urban greening monitoring. By acquiring remote sensing images using cameras, sensors, and other equipment, and combining them with computer vision and artificial intelligence models, it enables automated assessment of vegetation status.

[0003] Existing patent CN117292282A discloses a method and system for monitoring the growth of greenery in urban landscapes based on high-resolution UAV remote sensing. It acquires image data via UAV, processes it to extract morphological, textural, and color features of vegetation, and uses a ConvLSTM model to construct a growth analysis model, thus improving the convenience and efficiency of monitoring to some extent. However, this patent still has significant technical shortcomings, with the core issues being incomplete monitoring and a lack of aesthetic assessment due to canopy shading. In urban landscaping, dense areas of tall trees are common, and their dense canopies completely obscure the lower shrubs, ground cover, and other low-lying vegetation. The single-view camera or conventional remote sensing methods used in this patent cannot penetrate the canopy to obtain data on the underlying vegetation; they can only assess growth based on canopy surface information. More importantly, the patent does not consider one of the core requirements of landscaping – visual aesthetics assessment. Aesthetic indicators such as the regularity of vegetation morphology (e.g., symmetry, uniformity), color consistency, and layer coordination directly affect citizens' visual experience on the ground. Existing methods cannot reproduce the morphology of understory vegetation, resulting in a serious mismatch between monitoring results and actual experience, making it difficult to support refined landscaping management and landscape optimization decisions. Summary of the Invention

[0004] This application provides a method and system for monitoring the status of landscaping based on UAV remote sensing, which solves the problems in the prior art that rely on vegetation canopy data but cannot obtain data on the underlying vegetation and cannot assess the visual aesthetics of the underlying vegetation.

[0005] On the one hand, embodiments of this application provide a method for monitoring the status of landscaping based on unmanned aerial vehicle (UAV) remote sensing, including:

[0006] The drone was used to collect multi-view dual-band fused point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique views.

[0007] The multi-view dual-band fused point cloud is registered to the reference coordinate system. The transformation matrix is ​​iteratively optimized using ground feature points as registration anchor points to obtain a multi-view merged point cloud under a unified coordinate system. The holes in the multi-view merged point cloud are filled to generate a preliminary high-density LiDAR point cloud.

[0008] The preliminary high-density LiDAR point cloud is layered according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover to form a layered preliminary high-density LiDAR point cloud.

[0009] The absolute blind zone in the preliminary high-density LiDAR point cloud after layering is identified. The pixels in the absolute blind zone of the multi-view hyperspectral image are decomposed into pixels and the forest understory vegetation abundance is calculated. Based on the forest understory vegetation abundance, the blind zone with effective vegetation is determined.

[0010] The morphological parameters of vegetation in blind areas with effective vegetation are inferred. Virtual point clouds are generated based on the morphological parameters. The virtual point clouds and the layered preliminary high-density LiDAR point clouds are stitched together to obtain a complete high-density LiDAR point cloud.

[0011] Extract matching feature point pairs from multi-view hyperspectral images from different perspectives, generate SfM sparse point clouds based on matching feature point pairs, fuse the SfM sparse point clouds with complete high-density LiDAR point clouds to form fused point clouds, and establish a complete three-dimensional mesh model based on fused point clouds.

[0012] Extract color information from multi-view hyperspectral images and map the color information onto a complete 3D mesh model to form a textured complete 3D model;

[0013] Multi-dimensional features are extracted from textured complete 3D models and multi-view hyperspectral images, and input into an improved ConvLSTM model to obtain a full-area vegetation feature set including growth quantification scores and aesthetic features.

[0014] Growth and aesthetics assessments are conducted based on a set of vegetation characteristics across the entire region.

[0015] On the other hand, embodiments of this application also provide a landscaping status monitoring system based on UAV remote sensing, including:

[0016] The drone is used to acquire multi-view dual-band fusion point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique viewpoints.

[0017] The preliminary point cloud generation module is used to register the multi-view dual-band fused point cloud to the reference coordinate system. It iteratively optimizes the transformation matrix with ground feature points as registration anchor points to obtain the multi-view merged point cloud under the unified coordinate system, fills the holes in the multi-view merged point cloud, and generates a preliminary high-density LiDAR point cloud.

[0018] The point cloud layering module is used to layer the preliminary high-density LiDAR point cloud according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover, forming a layered preliminary high-density LiDAR point cloud.

[0019] The blind zone identification module is used to identify absolute blind zones in the preliminary high-density LiDAR point cloud after layering. It performs pixel decomposition on pixels in the absolute blind zone of the multi-view hyperspectral image, calculates the abundance of understory vegetation, and determines the blind zone with effective vegetation based on the abundance of understory vegetation.

[0020] The point cloud stitching module is used to infer the morphological parameters of vegetation in blind areas where effective vegetation exists, generate virtual point clouds based on morphological parameters, and stitch the virtual point clouds with the layered preliminary high-density LiDAR point clouds to obtain a complete high-density LiDAR point cloud.

[0021] The mesh building module is used to extract matching feature point pairs from multi-view hyperspectral images from different perspectives, generate SfM sparse point clouds based on the matching feature point pairs, fuse the SfM sparse point clouds with the complete high-density LiDAR point clouds to form fused point clouds, and build a complete 3D mesh model based on the fused point clouds.

[0022] The texture mapping module is used to extract color information from multi-view hyperspectral images and map the color information onto a complete 3D mesh model to form a textured complete 3D model.

[0023] The feature set building module is used to extract multi-dimensional features from textured full 3D models and multi-view hyperspectral images, input them into the improved ConvLSTM model, and obtain a full-area vegetation feature set including growth quantification scores and aesthetic features.

[0024] The status assessment module is used to assess growth and aesthetics based on the vegetation feature set of the entire region.

[0025] The method and system for monitoring the status of landscaping based on UAV remote sensing disclosed in this application have the following advantages:

[0026] 1. Strong canopy penetration and data completion capabilities, with no blind spots. Through dual-band LiDAR multi-view fusion, hyperspectral hybrid pixel decomposition, and virtual point cloud completion technology, the data loss rate of understory vegetation is reduced to below 7.8%, completely solving the blind spot problem caused by canopy shading in densely wooded areas.

[0027] 2. Accurate 3D morphology and texture reproduction to support aesthetic assessment. Through point cloud fusion, Poisson surface reconstruction, and hyperspectral texture mapping, a complete 3D model with realistic texture is generated, accurately reproducing morphological details such as vegetation height, crown width, and symmetry, with a morphological reproduction error (MAE) ≤0.12m. Quantifying core aesthetic indicators such as symmetry and color consistency fills the gap in existing technologies that cannot assess the visual aesthetics of vegetation, meeting the needs related to ground visual perception.

[0028] 3. Comprehensive evaluation dimensions, highly aligned with actual experience. A dual-dimensional evaluation system considering both growth and aesthetics is constructed. The improved ConvLSTM model output has a correlation coefficient of 0.86 with the visual scores from ground volunteers, and the overall evaluation shows an 89.3% agreement rate with on-site assessments by professionals. This avoids the problem of existing patents focusing only on growth and being disconnected from ground experience; the evaluation results are more closely aligned with the actual needs of landscape and green space management.

[0029] 4. The process is closed-loop, practical, and efficient, requiring no ground data support. It relies entirely on multi-source data collection and processing by drones, eliminating the need for additional on-site investigations. The final output includes a 3D visualization map, heat map, and rectification suggestion report, directly providing decision support for refined landscape management (replanting, pruning, water and fertilizer control). It balances monitoring efficiency with practical applicability, offering convenient operation and controllable costs. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart illustrating a method for monitoring the status of landscaping based on unmanned aerial vehicle (UAV) remote sensing, provided as an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Figure 1 A flowchart illustrating a method for monitoring the status of urban greening based on unmanned aerial vehicle (UAV) remote sensing, provided in this application embodiment. This application embodiment provides a method for monitoring the status of urban greening based on UAV remote sensing, including:

[0034] The S100 uses drones to collect multi-view dual-band fusion point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique angles.

[0035] For example, before collecting data, it is necessary to first clarify the boundaries and vegetation distribution characteristics of the monitoring area, and then divide it into sub-areas such as densely wooded areas, mixed vegetation areas, and open grasslands to determine the key monitoring areas. For example, the canopy closure of densely wooded areas should be ≥0.7. Based on the payload capacity of the UAV, complete the installation and debugging of the dual-band LiDAR, hyperspectral camera, and high-precision positioning module to ensure that the equipment synchronization meets the data acquisition requirements.

[0036] Then, the flight was executed according to the main flight path plus supplementary flight paths. The main flight path adopted a 0° direct view flight, with a flight altitude of 40-60m, a forward overlap rate of ≥85%, and a lateral overlap rate of ≥70%. For densely wooded areas, three supplementary flight paths with oblique perspectives of 15°, 30°, and 60° were added. Each perspective flight path was parallel to the main flight path, with a forward overlap rate of ≥80% and a lateral overlap rate of ≥60%. During the flight, dual-band LiDAR and hyperspectral cameras were simultaneously triggered to acquire data, ensuring that each frame of hyperspectral imagery corresponded to a set of LiDAR point cloud data.

[0037] The output multi-view dual-band LiDAR raw data includes point clouds in the 1550nm and 905nm bands from each viewpoint, containing three-dimensional coordinates, reflectance intensity, and band labels. The multi-view hyperspectral raw data contains 50 bands, covering the 400-1000nm spectral range. These data cover the entire monitoring area, and the multi-source data are spatiotemporally synchronized.

[0038] Furthermore, after acquiring multi-view dual-band fused point clouds and multi-view hyperspectral images, preprocessing is performed. The preprocessing of multi-view dual-band fused point clouds includes format conversion, noise removal, and multi-view dual-band fusion. The preprocessing of multi-view hyperspectral images includes radiometric correction, atmospheric correction, registration and stitching, and cropping.

[0039] Specifically, the format conversion involves batch converting multi-view dual-band LiDAR raw data (.las format) into the .pcd format supported by the Point Cloud Library (PCL), and extracting the three-dimensional coordinates (X,Y,Z), reflection intensity, band identification, and view identification information of each point.

[0040] Noise removal employs an adaptive threshold filtering algorithm, setting thresholds for the number of local neighborhood points and reflection intensity, with a threshold of ≥0.1 for the 1550nm band and ≥0.15 for the 905nm band. This removes invalid data such as isolated points caused by drone jitter and bird interference points, while retaining effective vegetation and ground point clouds.

[0041] Split-view dual-band fusion is a process that combines point cloud data from each viewpoint with point clouds in the 1550nm and 905nm bands, weighted by reflection intensity. The weight of the 1550nm point cloud is set to 0.6, and the weight of the 905nm point cloud is set to 0.4, to generate a single-view dual-band fused point cloud.

[0042] Radiometric correction is based on the calibration gain coefficient K and offset coefficient B provided by the camera manufacturer. The original DN value of the hyperspectral image is converted into a radiance value L using the formula L=DN×K+B, thereby correcting the sensor response inhomogeneity error.

[0043] Atmospheric correction uses the FLAASH atmospheric correction model, inputting meteorological data (temperature, humidity, air pressure) and sensor parameters at the time of acquisition, selecting a rural atmospheric model and a continental aerosol type, retrieving surface reflectance, and eliminating atmospheric scattering and absorption interference.

[0044] Registration and stitching are based on 0° view hyperspectral images. The SIFT feature point matching algorithm and RANSAC mismatch removal algorithm are used to complete the registration of multi-view hyperspectral images. Then, the linear fusion algorithm is used to stitch the registered images, process the brightness difference at the stitching edge, and generate a full-area hyperspectral image.

[0045] Cropping involves cropping and stitching the hyperspectral image based on the boundary vector file of the monitoring area, removing invalid background areas, and retaining the valid monitoring range data.

[0046] The multi-view dual-band fused point cloud output from this step contains four independent point clouds, corresponding to 0°, 15°, 30°, and 60° viewpoints, respectively. The preprocessed multi-view hyperspectral imagery includes registered single-view images and full-area mosaic cropped images. Both types of data retain viewpoint identifiers and geographic location information, which can be correlated via coordinates.

[0047] S110 registers the multi-view dual-band fused point cloud to the reference coordinate system, uses ground feature points as registration anchors to iteratively optimize the transformation matrix, obtains the multi-view merged point cloud under a unified coordinate system, fills the holes in the multi-view merged point cloud, and generates a preliminary high-density LiDAR point cloud.

[0048] For example, the 0° view dual-band fused point cloud is selected as the registration reference because it has complete coverage and strong stability of the frontal view data.

[0049] The Iterative Closest Point (ICP) algorithm was used to register dual-band fused point clouds from 15°, 30°, and 60° perspectives to the reference coordinate system. Ground feature points (road edges, flower bed boundaries) were used as registration anchor points, and the transformation matrix was iteratively optimized to ensure that the registration error was ≤2cm, finally obtaining a multi-view merged point cloud in a unified coordinate system.

[0050] The multi-view merged point cloud is deduplicated by removing points with ≥95% overlap in 3D coordinates. Then, the Kriging interpolation algorithm is used to fill in the tiny point cloud holes in the non-blind areas, with an interpolation radius of 0.5m and a minimum of ≥3 interpolation points, generating a preliminary high-density LiDAR point cloud with a density ≥150 points / m². 2 .

[0051] S120, the preliminary high-density LiDAR point cloud is layered according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover to form a layered preliminary high-density LiDAR point cloud.

[0052] For example, progressive morphological filtering (PMF) is used, with an initial window size of 3m, an increment step of 2m, and a maximum window size of 10m. Ground points and non-ground points (vegetation points) are separated by morphological opening operations to generate a digital elevation model (DEM).

[0053] The actual height H of each point is calculated based on the DEM, where H = point cloud Z coordinate - corresponding DEM elevation. Then, the points are layered according to height thresholds: the upper layer (H≥3m) is the tree canopy point cloud, the middle layer (0.5m<H<3m) is the understory shrub point cloud, and the lower layer (0<H≤0.5m) is the understory ground cover point cloud. The preliminary high-density LiDAR point cloud after layering is output.

[0054] S130 identifies absolute blind zones in the preliminary high-density LiDAR point cloud after layering, performs pixel decomposition on pixels in the absolute blind zone of the multi-view hyperspectral image, calculates the abundance of understory vegetation, and determines the blind zone with effective vegetation based on the abundance of understory vegetation.

[0055] For example, by traversing the initial high-density LiDAR point cloud after layering, continuous point cloud-free regions (area ≥ 1m²) are identified. 2 The area is marked as an absolute blind zone, and its geographical location, boundary coordinates and surrounding vegetation type are recorded (based on the layered point cloud).

[0056] Pixels from the full-area hyperspectral image corresponding to the absolute blind zone were extracted. A linear hybrid pixel decomposition algorithm was used, and endmember spectra of tree leaves, shrub leaves, and garden soil from the USGS public spectral library were selected. The forest understory vegetation abundance f was calculated using the least squares method. If f ≥ 0.3, the blind zone was determined to have effective forest understory vegetation; if f < 0.1, it was determined to be bare soil or litter; if 0.1 ≤ f < 0.3, it was determined to be sparse vegetation.

[0057] S140, infer the morphological parameters of vegetation in the blind area where effective vegetation exists, generate a virtual point cloud based on the morphological parameters, and stitch the virtual point cloud and the layered preliminary high-density LiDAR point cloud to obtain a complete high-density LiDAR point cloud.

[0058] For example, for blind areas with effective vegetation, morphological parameters of the same type of understory vegetation within a 1-2m radius of the blind area are extracted, and the morphological parameters of the vegetation in the blind area are estimated by combining the ratio of the abundance of understory vegetation in the blind area to the abundance of understory vegetation in the surrounding area.

[0059] Specifically, for blind areas where effective vegetation is determined to exist, morphological parameters of the same type of understory vegetation within a 1-2m radius around the blind area are extracted, including average height H_avg, average crown width CW_avg, average symmetry S_avg, and average density D_avg. Combined with the ratio of the abundance f of the understory vegetation in the blind area to the abundance f0 of the surrounding vegetation, the morphological parameters of the vegetation in the blind area are calculated as follows: Blind area vegetation height H = H_avg × (f / f0), crown width CW = CW_avg × (f / f0), symmetry S = S_avg × (f / f0), and density D = D_avg × (f / f0).

[0060] Furthermore, based on the inferred morphological parameters of the vegetation in the blind zone, a virtual point cloud is generated within the three-dimensional space of the blind zone according to the distribution pattern of the vegetation point cloud. The virtual point cloud is then stitched together with the layered preliminary high-density LiDAR point cloud according to the coordinate system. The duplicate points after stitching are removed to form a complete high-density LiDAR point cloud without blind zones in the entire region.

[0061] Specifically, based on the inferred vegetation morphology parameters in the blind zone, a virtual point cloud is generated within the three-dimensional space of the blind zone, following the distribution pattern of the real vegetation point cloud (uniform density within the canopy and gradual change in height from bottom to top). The density is consistent with the preliminary high-density LiDAR point cloud (≥150 points / m²). 2 This ensures that the three-dimensional coordinates and density distribution of the virtual point cloud match the real vegetation characteristics.

[0062] The virtual point cloud is stitched together with the layered preliminary high-density LiDAR point cloud according to the coordinate system to cover the holes in the blind spot point cloud. At the same time, duplicate points after stitching are removed to finally form a complete point cloud without blind spots in the entire area.

[0063] The density of the complete high-density LiDAR point cloud output by this step is ≥150 points / m². 2 The point cloud contains layered information on tree canopy, understory shrubs, and understory ground cover. It includes real points in non-blind areas and virtual points in blind areas, covering the entire monitoring area. This point cloud carries the basic three-dimensional morphological information of vegetation in all areas, providing core data for subsequent point cloud fusion.

[0064] S150 extracts matching feature point pairs from multi-view hyperspectral images at different perspectives, generates SfM sparse point clouds based on the matching feature point pairs, fuses the SfM sparse point clouds with the complete high-density LiDAR point clouds to form fused point clouds, and establishes a complete three-dimensional mesh model based on the fused point clouds.

[0065] For example, the SIFT algorithm is used to extract feature points from each image of the preprocessed multi-view hyperspectral image, including key point coordinates and descriptors, to ensure that the feature points can cover key texture details of the vegetation surface.

[0066] Then, the FLANN matching algorithm is used to match feature points between images from different perspectives. The RANSAC algorithm is used to remove mismatched points, that is, to retain feature point pairs with a matching accuracy of ≥90%, so as to ensure that the matching results can reflect the spatial correspondence between images.

[0067] Furthermore, based on matched feature point pairs, bundle adjustment (BA) is used to optimize the camera's intrinsic parameters (focal length, principal point coordinates) and extrinsic parameters (position, attitude) to ensure that the camera parameters accurately reflect the spatial attitude during image capture. The optimized camera parameters and matched feature point pairs are used to triangulate and calculate 3D point coordinates, generating a preliminary SfM sparse point cloud containing 3D positional information corresponding to vegetation surface texture. Based on the capture position information of the multi-view hyperspectral images recorded by the high-precision positioning module, the preliminary SfM sparse point cloud is transformed to a coordinate system consistent with the complete high-density LiDAR point cloud, ensuring spatial alignment and forming the SfM sparse point cloud.

[0068] The SfM sparse point cloud contains three-dimensional location information of vegetation surface texture and only covers non-blind areas. It is consistent with the coordinate system of the complete high-density LiDAR point cloud. This point cloud will be used to supplement the detailed location information of vegetation in non-blind areas and improve the accuracy of subsequent point cloud fusion.

[0069] Since the complete high-density LiDAR point cloud and the SfM sparse point cloud have been generated based on the same coordinate system, only the following fine-tuning optimization is required: select ground feature points (such as road corners and flower bed edges) and vegetation feature points (such as tree canopy apex and shrub canopy edge) in the overlapping area (non-blind area) of the two as registration anchor points, and use the weighted iterative nearest point algorithm (W-ICP) to fine-tune the coordinates of the SfM sparse point cloud so that the point cloud deviation in the overlapping area of ​​the two is ≤2cm.

[0070] Furthermore, when fusing SfM sparse point clouds and complete high-density LiDAR point clouds, a weighted average method is used for fusion in overlapping areas, while data from virtual point clouds in the complete high-density LiDAR point cloud are retained for blind areas.

[0071] Specifically, for the overlapping regions outside the blind zone, a weighted average method is used to fuse the 3D coordinates of the complete high-density LiDAR point cloud and the SfM sparse point cloud. The weight of the complete high-density LiDAR point cloud is set to 0.7, and the weight of the SfM sparse point cloud is set to 0.3. The calculation formula is: P 融合 =0.7×P LiDAR +0.3×P SfM , where P 融合P represents the 3D coordinates of the merged point. LiDAR Let P be the coordinates of a LiDAR point. SfM Here are the coordinates of point SfM.

[0072] For blind areas, since there is no corresponding data in the SfM sparse point cloud, the virtual point cloud data in the complete high-density LiDAR point cloud is directly retained to ensure that the vegetation morphology information in the blind area is not lost.

[0073] The fused point cloud is denoised by removing isolated points (points with fewer than 3 local neighbors are considered isolated) and undergoes density equalization to generate the final fused point cloud. The density of this fused point cloud is ≥150 points / m². 2 The coordinate system consists of non-blind zone fusion points and blind zone virtual points, combining the high-precision 3D structure of LiDAR point clouds with the texture detail location information of SfM point clouds. This point cloud provides a complete and high-precision spatial data foundation for subsequent 3D mesh construction.

[0074] Furthermore, the K (K=12) nearest neighbor algorithm is used to find the neighboring points of each LiDAR original point and fused point in the fused point cloud, constructing a local plane. The normal vector is solved based on the plane equation, and its direction is perpendicular to the local plane, ensuring that the normal vector reflects the spatial pose of the vegetation surface. For the SfM supplementary points in the fused point cloud, the normal vector is corrected by linear interpolation of the normal vectors of neighboring LiDAR points, avoiding normal vector errors caused by the low density of the SfM point cloud and ensuring continuous and consistent normal vectors. The Laplace smoothing algorithm is used to globally adjust the normal vectors of all points, eliminating local normal vector conflicts, such as opposite directions of normal vectors between adjacent points, ensuring a continuous and uniform distribution of normal vectors. The fused point cloud with normal vectors is input into the Poisson surface reconstruction algorithm, with a reconstruction depth of 10 levels and a sampling density factor of 1.5. Surface fitting is performed on the spatial region enclosed by the fused point cloud to generate a complete 3D mesh model. During reconstruction, it is ensured that the mesh accurately reproduces the subtle morphology of vegetation, such as the direction of shrub branches and ground cover undulations, without obvious voids or jagged surfaces.

[0075] The generated 3D mesh model is then simplified and smoothed as follows: a quadrilateral mesh simplification algorithm is used to optimize the number of mesh faces to a reasonable range while preserving the core morphological features; a bilateral filtering algorithm is used to smooth the mesh surface to eliminate surface roughness problems generated during reconstruction while preserving vegetation edge details.

[0076] The output complete 3D mesh model contains a continuous and smooth 3D surface structure of vegetation, covering the entire monitoring area without data gaps, and can accurately reproduce the morphological details of trees, shrubs and ground cover. This model provides a stable structural framework for subsequent texture mapping.

[0077] S160 extracts color information from multi-view hyperspectral images and maps the color information onto a complete 3D mesh model to form a textured complete 3D model.

[0078] For example, based on preprocessed multi-view hyperspectral images and a complete 3D mesh model, the corresponding position of each face of the mesh on the hyperspectral image is calculated using the camera projection matrix to generate texture coordinates. For the mesh area covered by the multi-view images, the image with the highest texture clarity and the most uniform illumination is selected as the main texture source to ensure the clarity and consistency of texture mapping.

[0079] Specifically, RGB color information is extracted from the preprocessed hyperspectral image to generate a texture image, ensuring that the texture image color matches the true color of the vegetation. The extracted texture image is then mapped onto the surface of a complete 3D mesh model using texture coordinates. A bilinear interpolation algorithm is employed to optimize the texture mapping accuracy, avoiding distortion problems such as texture stretching and warping. For shaded areas under the forest canopy, textures from adjacent viewpoints are used to supplement the texture, ensuring the integrity of the texture in the shaded areas.

[0080] Furthermore, after mapping the color information to the complete 3D mesh model, color uniformity optimization is performed on the mapped complete 3D mesh model. Color uniformity optimization includes local color correction, spectral consistency verification, and global color balance.

[0081] Among them, local color correction is to calculate the color mean and standard deviation of the texture of the same vegetation patch, and use an adaptive histogram equalization algorithm to adjust the brightness and contrast so that the color uniformity within the patch meets the requirement of standard deviation ≤ 20.

[0082] Spectral consistency verification combines the reflectance of the red edge band of hyperspectral images to correct color values. Healthy vegetation (with significant peaks in red edge reflectance) corresponds to bright green texture, while poorly growing vegetation (with low red edge reflectance) corresponds to light green or yellow texture, ensuring that the texture color is both visually realistic and reflects the health status of the vegetation.

[0083] Global color balancing is a process that balances the texture of a 3D mesh model across the entire region, eliminating color deviations caused by different viewing angles and lighting conditions, and ensuring that the texture color transitions naturally and consistently across the entire region.

[0084] The textured, complete 3D model output by this step contains accurate 3D morphological structure and realistic, uniform hyperspectral texture color, which can intuitively reflect the morphological characteristics and health status of vegetation in the entire area. This model provides core visualization data support for subsequent vegetation feature extraction and aesthetic assessment.

[0085] S170 extracts multi-dimensional features from a textured complete 3D model and multi-view hyperspectral images, inputs them into an improved ConvLSTM model, and obtains a full-area vegetation feature set including growth quantification scores and aesthetic features.

[0086] For example, multi-dimensional features are extracted from a textured complete 3D model and preprocessed hyperspectral imagery to form an original feature set. These multi-dimensional features include morphological features, spectral features, and occlusion and health features. Morphological features are extracted from the 3D model, including vegetation height, crown width, symmetry, density uniformity, and hierarchical coordination (height differences between trees, shrubs, and ground cover), totaling five indicators. Spectral features are extracted from the hyperspectral imagery, including NDVI (Normalized Difference Vegetation Index), red edge position, red edge amplitude, color mean, and color standard deviation, totaling five indicators. Occlusion and health features are extracted from the hyperspectral imagery and the 3D model, including occlusion coefficient, understory vegetation abundance, and percentage of withered / yellowed area, totaling three indicators.

[0087] The original feature set is standardized by mapping the feature values ​​to the 0-1 interval to eliminate dimensional differences and generate a standardized feature set.

[0088] Furthermore, the improved ConvLSTM model includes an input layer, a convolutional layer, a ConvLSTM hidden layer, a feature fusion layer, and an output layer. The output layer includes a growth feature output layer and an aesthetic feature output layer.

[0089] Specifically, the input layer receives a standardized feature set with the following dimensions: batch size, time step, and feature dimension. The feature dimension includes morphological features, spectral features, and occlusion and health features, with a total feature dimension of 13.

[0090] The convolutional layer contains two convolutional blocks. Each convolutional block consists of one sub-convolutional layer, one batch normalization (BN) layer, and one ReLU activation function. The kernel size of the sub-convolutional layer is 3×3, the stride is 1, and the padding method is the same. The first convolutional block has 32 output channels, and the second has 64, which are used to extract local features.

[0091] The ConvLSTM hidden layer consists of three ConvLSTM units, each with a hidden layer dimension of 128. The forget gate, input gate, and output gate all use the sigmoid activation function, while the cell state update uses the tanh activation function. Residual connections are used between units to enhance the ability to learn long-term dependencies and alleviate the gradient vanishing problem.

[0092] The feature fusion layer employs an attention mechanism to assign weights to the feature maps output by the ConvLSTM hidden layer, focusing on features related to shape and aesthetics, such as symmetry and color uniformity. The fused feature dimension is 128.

[0093] The output layer is divided into two branches. The growth feature output layer contains one fully connected layer with an output dimension of 3, corresponding to the growth levels: good, medium, and poor. The aesthetic feature output layer contains one fully connected layer with an output dimension of 5, and the corresponding aesthetic indicators are symmetry, uniformity, color consistency, hierarchical coordination, and the proportion of no bad conditions.

[0094] In the process of improving the training of the ConvLSTM model, the training dataset was selected from three different types of urban greening areas, including urban parks, residential green spaces and road green belts. Each of the 1,000 sets of sample data collected included UAV multi-source data (LiDAR + hyperspectral) and ground-measured data (vegetation morphology parameters, growth level, and aesthetic score).

[0095] After collecting the data samples, five landscape professionals were invited to label the samples. The growth level was labeled based on the health status of the vegetation (leaf color, withering rate), and the aesthetic index was rated on a scale of 1 to 5 (5 being the best). The average value was used as the label. Then, the dataset was divided into a training set (700 sets) and a test set (300 sets) in a 7:3 ratio. The training set used 5-fold cross-validation to avoid overfitting.

[0096] The loss function used during training is a joint loss function that takes into account both growth classification and aesthetic regression tasks. The loss function is expressed as L = α × L1 + β × L2, where L1 is the cross-entropy loss for growth classification, L2 is the mean squared error (MSE) loss for aesthetic regression, α = 0.4, and β = 0.6.

[0097] The optimizer used is the Adam optimizer, with an initial learning rate of 0.001. The learning rate decay strategy is cosine annealing decay, which decays once every 10 epochs, with a minimum learning rate of 0.0001.

[0098] Batch size = 32, training iterations = 100 epochs, early stopping strategy (patience = 15) is adopted, that is, if the loss on the test set does not decrease for 15 consecutive epochs, then training is stopped and the optimal model is saved.

[0099] Regularization uses L2 regularization (weight decay coefficient = 0.0001) and a Dropout layer (dropout rate = 0.2) to prevent model overfitting.

[0100] The model training process is as follows: input the preprocessed training data into the model, train according to the above parameters, monitor the loss changes of the training set and validation set in real time, and adjust the learning rate and regularization parameters. After training, evaluate the model performance using the test set. The growth classification accuracy is ≥88%, and the aesthetic regression error (MAE) is ≤0.3, meeting the monitoring requirements.

[0101] After improving the ConvLSTM model to output the full-area vegetation feature set, post-processing is required. This post-processing includes: converting the growth level into a quantitative score, where good = 9 points, medium = 6 points, and poor = 3 points; performing consistency verification on aesthetic features and removing outliers, i.e., feature values ​​that exceed the 0-1 range or conflict with the intuitive features of the 3D model, to ensure that the feature data is accurate and reliable.

[0102] The vegetation feature set output by this step includes the growth quantification score of each vegetation patch and 5 aesthetic normalized feature values, covering the entire monitoring area. This feature set provides direct data input for the subsequent comprehensive assessment of the status of landscaping.

[0103] S180 assesses growth and aesthetics based on a set of vegetation characteristics across the entire region.

[0104] For example, a comprehensive growth score is calculated based on the quantitative growth score and NDVI value of the vegetation feature set of the whole region. The calculation process is as follows: the quantitative growth score is the base score. When NDVI≥0.7, the base score is increased by 1 point. When 0.5≤NDVI<0.7, the base score remains unchanged. When NDVI<0.5, the base score is decreased by 1 point. When the red edge position shifts to red, 0.5 points are added and when it shifts to blue, 0.5 points are subtracted. The final score is the comprehensive growth score.

[0105] The overall aesthetic score is calculated using a weighted summation method, with the following weights: symmetry 0.25, uniformity 0.2, color consistency 0.25, hierarchical coordination 0.15, and percentage of no defective features 0.15. The formula is: Overall Aesthetic Score = (Symmetry × 0.25 + Uniformity × 0.2 + Color Consistency × 0.25 + Hierarchical Coordination × 0.15 + Percentage of No Defective Features × 0.15) × 10. Then, based on the overall aesthetic score, a grade is assigned: ≥8 points is "Excellent," 6-7.9 points is "Good," 4-5.9 points is "Average," and <4 points is "Poor."

[0106] After determining the overall growth score and the overall aesthetic score, a weighted summation method is used to calculate the overall status score. The overall growth score has a weight of 0.4, and the overall aesthetic score has a weight of 0.6. The calculation formula is: Overall Status Score = Overall Growth Score × 0.4 + Overall Aesthetic Score × 0.6. The overall rating is divided as follows: ≥8 points = "Excellent", 6-7.9 points = "Good", 4-5.9 points = "Satisfactory", and <4 points = "Requires Rectification".

[0107] After obtaining the comprehensive status score, based on the textured complete 3D model, a 3D visualization map supporting aerial, ground, and layered perspectives will be generated. These maps can mark the comprehensive level and key feature shortcomings of each area. At the same time, using the vector map of the monitored area as the base map, different colors (red = requires rectification, yellow = qualified, green = good, dark green = excellent) are used to mark the comprehensive score level of each area, intuitively displaying the status distribution.

[0108] Finally, a detailed assessment report will be output, including an overview of the monitoring area, statistics of growth and aesthetic assessment results, distribution of areas at each level, details of blind spot assessment, and targeted rectification suggestions (such as replanting in bare soil areas, pruning asymmetrical shrubs, and water and fertilizer management for withered vegetation).

[0109] This application also provides a landscaping status monitoring system based on UAV remote sensing, including:

[0110] The drone is used to acquire multi-view dual-band fusion point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique viewpoints.

[0111] The preliminary point cloud generation module is used to register the multi-view dual-band fused point cloud to the reference coordinate system. It iteratively optimizes the transformation matrix with ground feature points as registration anchor points to obtain the multi-view merged point cloud under the unified coordinate system, fills the holes in the multi-view merged point cloud, and generates a preliminary high-density LiDAR point cloud.

[0112] The point cloud layering module is used to layer the preliminary high-density LiDAR point cloud according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover, forming a layered preliminary high-density LiDAR point cloud.

[0113] The blind zone identification module is used to identify absolute blind zones in the preliminary high-density LiDAR point cloud after layering. It performs pixel decomposition on pixels in the absolute blind zone of the multi-view hyperspectral image, calculates the abundance of understory vegetation, and determines the blind zone with effective vegetation based on the abundance of understory vegetation.

[0114] The point cloud stitching module is used to infer the morphological parameters of vegetation in blind areas where effective vegetation exists, generate virtual point clouds based on morphological parameters, and stitch the virtual point clouds with the layered preliminary high-density LiDAR point clouds to obtain a complete high-density LiDAR point cloud.

[0115] The mesh building module is used to extract matching feature point pairs from multi-view hyperspectral images from different perspectives, generate SfM sparse point clouds based on the matching feature point pairs, fuse the SfM sparse point clouds with the complete high-density LiDAR point clouds to form fused point clouds, and build a complete 3D mesh model based on the fused point clouds.

[0116] The texture mapping module is used to extract color information from multi-view hyperspectral images and map the color information onto a complete 3D mesh model to form a textured complete 3D model.

[0117] The feature set building module is used to extract multi-dimensional features from textured full 3D models and multi-view hyperspectral images, input them into the improved ConvLSTM model, and obtain a full-area vegetation feature set including growth quantification scores and aesthetic features.

[0118] The status assessment module is used to assess growth and aesthetics based on the vegetation feature set of the entire region.

[0119] Experimental process

[0120] 1. Experimental design.

[0121] To verify the effectiveness of this method, a city park with an area of ​​approximately 5000 m² was selected as the experimental area. 2 The area includes typical landscaping scenes such as densely wooded areas, mixed vegetation areas (trees + shrubs + ground cover), and open grasslands. The densely wooded area (approximately 1500 square meters) is a prime example. 2 The key testing area was identified as having a canopy closure of ≥0.8, indicating significant canopy shading. The experiment used existing patent CN117292282A as a comparison object, evaluating it from four dimensions: data integrity, morphological accuracy, aesthetic assessment capability, and consistency with ground perception.

[0122] 1.1 Experimental equipment.

[0123] This method uses a DJI M300 drone equipped with a 16-line dual-band LiDAR (1550nm + 905nm), a miniature hyperspectral camera (50 bands), and an RTK positioning module. The comparison method uses the same drone equipped with a standard RGB camera.

[0124] 1.2 Evaluation Indicators.

[0125] Data completeness: Calculate the missing rate of understory vegetation data in densely wooded areas, i.e., the area of ​​areas without valid data / the total area.

[0126] Morphological reconstruction accuracy: Twenty understory shrub samples were selected, and the measured morphological parameters (height, crown width, symmetry) were compared with the inversion results of the two methods to calculate the mean absolute error (MAE).

[0127] Aesthetic assessment accuracy: Ten volunteers were invited to rate the aesthetic appeal of the ground in densely wooded areas (1-10 points). The correlation coefficients between the aesthetic scores of the two methods and the volunteer scores were compared, specifically using the Pearson correlation coefficient.

[0128] Overall assessment consistency: Compare the overall scores of the two methods with the results of on-site assessments by landscape professionals (consistency rate = area of ​​consistent scores / total area).

[0129] 2. Experimental process and results.

[0130] 2.1 Data Acquisition and Processing.

[0131] The flight plan according to this method was used to acquire dual-band LiDAR and hyperspectral data, and preprocessing, 3D structure extraction, texture reconstruction, and model inference were completed step by step. In contrast, the method used in comparison was to acquire RGB images according to the workflow of CN117292282A, and then perform data processing and growth analysis.

[0132] 2.2 Result Comparison and Analysis.

[0133] Data completeness: The missing data rate for forest understory vegetation in this method was 7.8%, with only a small number of extremely small areas (<0.5m²) missing. 2 Absolute blind zone. The missing rate of the comparative method was 35.2%, and about 1 / 3 of the densely wooded area had no understory vegetation data, which fully demonstrates that the canopy penetration ability of this method is significantly better than that of existing patents.

[0134] Morphological reconstruction accuracy: The MAE of this method for shrub height, crown width, and symmetry are 0.12m, 0.08m, and 0.05, respectively, showing small morphological reconstruction error. The comparative method, unable to obtain understory vegetation morphological data, can only infer from the canopy surface and cannot calculate morphological parameters, thus lacking any morphological reconstruction capability.

[0135] Aesthetic assessment accuracy: The Pearson correlation coefficient between the aesthetic score and the volunteer score of this method is 0.86, indicating a very strong correlation. The comparison method lacks an aesthetic assessment module and cannot output relevant results, while the correlation coefficient between the growth assessment results of the existing patent and the volunteer score is only 0.42, showing a disconnect from ground visual perception.

[0136] Overall assessment consistency: The overall score of this method matches the on-site assessment by professionals with a rate of 89.3%, significantly higher than the 62.5% of the comparative method. Due to the lack of understory data, the comparative method generally overestimates the overall assessment of densely wooded areas, while this method accurately reflects the morphological defects and aesthetic shortcomings of the understory vegetation, and the assessment results are more in line with reality.

[0137] 3. Experimental conclusions.

[0138] Experimental results show that this method, through dual-band LiDAR + hyperspectral multi-view fusion technology, successfully solves the data loss problem in the canopy-shaded area of ​​the existing patent CN117292282A, achieving accurate restoration of the three-dimensional morphology of understory vegetation. By adding an aesthetic assessment function through an improved ConvLSTM model, it fills a gap in existing technology. The monitoring results of this method are significantly superior to existing patents in terms of data completeness, morphological restoration accuracy, aesthetic assessment accuracy, and consistency with ground perception, providing more comprehensive and accurate decision support for refined management and landscape optimization in urban greening.

[0139] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0140] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for monitoring the status of garden greening based on unmanned aerial vehicle (UAV) remote sensing, characterized in that, include: The drone was used to collect multi-view dual-band fused point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique views. The multi-view dual-band fused point cloud is registered to the reference coordinate system, and the transformation matrix is ​​iteratively optimized with ground feature points as registration anchor points to obtain a multi-view merged point cloud under a unified coordinate system. The holes in the multi-view merged point cloud are filled to generate a preliminary high-density LiDAR point cloud. The preliminary high-density LiDAR point cloud is layered according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover to form the layered preliminary high-density LiDAR point cloud. The absolute blind zone in the preliminary high-density LiDAR point cloud after layering is identified. The pixels in the multi-view hyperspectral image located in the absolute blind zone are decomposed into pixels. The endmember spectra of tree leaves, shrub leaves and garden soil in the USGS public spectral library are selected. The forest understory vegetation abundance is calculated by the least squares method. The blind zone with effective vegetation is determined based on the forest understory vegetation abundance. The morphological parameters of vegetation in blind areas with effective vegetation are inferred, a virtual point cloud is generated based on the morphological parameters, and the virtual point cloud and the layered preliminary high-density LiDAR point cloud are stitched together to obtain a complete high-density LiDAR point cloud. Extract matching feature point pairs from the multi-view hyperspectral images at different viewpoints, generate SfM sparse point cloud based on the matching feature point pairs, fuse the SfM sparse point cloud with the complete high-density LiDAR point cloud to form a fused point cloud, and establish a complete three-dimensional mesh model based on the fused point cloud. The color information of the multi-view hyperspectral image is extracted and mapped onto the complete 3D mesh model to form a textured complete 3D model. Multi-dimensional features are extracted from the textured complete 3D model and the multi-view hyperspectral images, and input into the improved ConvLSTM model to obtain a full-area vegetation feature set including growth quantification scores and aesthetic features. Growth and aesthetics are assessed based on the vegetation feature set of the entire region.

2. The method for monitoring the status of garden greening based on UAV remote sensing according to claim 1, characterized in that, After acquiring the multi-view dual-band fused point cloud and the multi-view hyperspectral image, preprocessing is performed. The preprocessing of the multi-view dual-band fused point cloud includes format conversion, noise removal, and multi-view dual-band fusion. The preprocessing of the multi-view hyperspectral image includes radiometric correction, atmospheric correction, registration and stitching, and cropping.

3. The method for monitoring the status of garden greening based on UAV remote sensing according to claim 1, characterized in that, For blind areas with effective vegetation, the morphological parameters of the same type of understory vegetation within a 1-2m radius of the blind area are extracted, and the morphological parameters of the vegetation in the blind area are estimated by combining the ratio of the abundance of understory vegetation in the blind area to the abundance of understory vegetation in the surrounding area.

4. The method for monitoring the status of garden greening based on UAV remote sensing according to claim 1, characterized in that, Based on the inferred morphological parameters of the vegetation in the blind zone, a virtual point cloud is generated within the three-dimensional space of the blind zone according to the distribution pattern of the vegetation point cloud. The virtual point cloud is then stitched together with the layered preliminary high-density LiDAR point cloud according to the coordinate system. Duplicate points are removed after stitching to form a complete high-density LiDAR point cloud with no blind zones in the entire area.

5. A method for monitoring the status of landscaping based on UAV remote sensing according to claim 1, characterized in that, Based on the matched feature point pairs, the camera intrinsic and extrinsic parameters are optimized by bundle adjustment. The optimized camera parameters and the matched feature point pairs are used to triangulate and calculate the three-dimensional point coordinates to generate a preliminary SfM sparse point cloud. Based on the shooting location information of the multi-view hyperspectral image, the preliminary SfM sparse point cloud is transformed to a coordinate system consistent with the complete high-density LiDAR point cloud to form the SfM sparse point cloud.

6. A method for monitoring the status of landscaping based on UAV remote sensing according to claim 1, characterized in that, When fusing the SfM sparse point cloud and the complete high-density LiDAR point cloud, a weighted average method is used for fusing the overlapping areas, while the data of the virtual point cloud in the complete high-density LiDAR point cloud is retained for the blind areas.

7. A method for monitoring the status of landscaping based on UAV remote sensing according to claim 6, characterized in that, The K-nearest neighbor algorithm is used to find the neighboring points of each LiDAR original point and the fused point in the fused point cloud, a local plane is constructed, and the normal vector is solved based on the plane equation; the normal vector of the SfM supplementary point in the fused point cloud is corrected by linear interpolation of the normal vector of the neighboring LiDAR point. The normal vectors of all points are globally adjusted using the Laplacian smoothing algorithm. The fused point cloud with normal vectors is then input into the Poisson surface reconstruction algorithm to perform surface fitting on the spatial region enclosed by the fused point cloud, thereby generating the complete three-dimensional mesh model.

8. A method for monitoring the status of urban greening based on UAV remote sensing according to claim 1, characterized in that, After mapping the color information to the complete 3D mesh model, color uniformity optimization is performed on the mapped complete 3D mesh model. The color uniformity optimization includes local color correction, spectral consistency verification, and global color balance.

9. A method for monitoring the status of garden greening based on UAV remote sensing according to claim 1, characterized in that, The improved ConvLSTM model includes an input layer, a convolutional layer, a ConvLSTM hidden layer, a feature fusion layer, and an output layer. The output layer includes a growth feature output layer and an aesthetic feature output layer.

10. A system for monitoring the status of urban greening based on UAV remote sensing as described in any one of claims 1-9, characterized in that, include: The drone is used to acquire multi-view dual-band fusion point cloud and multi-view hyperspectral images of the garden from different perspectives, including the frontal view and multiple oblique viewpoints. The preliminary point cloud generation module is used to register the multi-view dual-band fused point cloud to the reference coordinate system, iteratively optimize the transformation matrix with ground feature points as registration anchor points, obtain the multi-view merged point cloud under the unified coordinate system, fill the holes in the multi-view merged point cloud, and generate a preliminary high-density LiDAR point cloud. The point cloud layering module is used to layer the preliminary high-density LiDAR point cloud according to the pre-set height thresholds of tree canopy, understory shrubs and understory ground cover, to form the layered preliminary high-density LiDAR point cloud. The blind spot identification module is used to identify absolute blind spots in the preliminary high-density LiDAR point cloud after layering, perform pixel decomposition on the pixels in the absolute blind spots of the multi-view hyperspectral image, calculate the abundance of understory vegetation, and determine the blind spots where effective vegetation exists based on the abundance of understory vegetation. The point cloud stitching module is used to infer the morphological parameters of vegetation in blind areas where effective vegetation exists, generate a virtual point cloud based on the morphological parameters, and stitch the virtual point cloud with the layered preliminary high-density LiDAR point cloud to obtain a complete high-density LiDAR point cloud. The mesh building module is used to extract matching feature point pairs of the multi-view hyperspectral images from different perspectives, generate SfM sparse point clouds based on the matching feature point pairs, fuse the SfM sparse point clouds and the complete high-density LiDAR point clouds to form a fused point cloud, and build a complete three-dimensional mesh model based on the fused point cloud. The texture mapping module is used to extract the color information of the multi-view hyperspectral image and map the color information onto the complete three-dimensional mesh model to form a textured complete three-dimensional model. The feature set building module is used to extract multi-dimensional features from the textured complete 3D model and the multi-view hyperspectral image, input them into the improved ConvLSTM model, and obtain a full-area vegetation feature set including growth quantification score and aesthetic features. The status assessment module is used to assess growth and aesthetics based on the vegetation feature set of the entire region.

Images