Landscape garden surveying and mapping method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a landscape garden surveying and mapping method based on artificial intelligence, and the method comprises the steps: 1, carrying an RGB camera, a laser radar and a multispectral sensor through an unmanned plane, collecting the multi-modal data of a landscape garden, and supplementing the near-ground point cloud data through a ground mobile robot; step 2, carrying out space-time reference alignment on the multi-modal data acquired in the step 1 by adopting an ICP-SIFT mixed registration algorithm to generate registered multi-period data; and step 3, constructing a space-time-spectrum tensor based on a registration result in the step 2. The technical scheme of combining an ICP-SIFT mixed registration technology and hypergraph constraint tensor decomposition is adopted, the technical effect of sub-centimeter-level multi-modal data space-time alignment and high-order correlation fusion is achieved, and compared with the defects that in the prior art, a single registration algorithm is prone to noise interference, and tensor decomposition neglects ecological correlation, the method has the advantage that the method has the advantage of being high in adaptability. The problems of multi-source data registration error accumulation and cross-modal feature fusion distortion are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and specifically provides a landscape garden surveying and mapping method based on artificial intelligence. Background Art

[0002] As a core link in the construction of smart cities and ecological planning, landscape garden surveying and mapping needs to integrate multi-modal perception data to achieve high-precision 3D modeling and dynamic monitoring. However, traditional surveying and mapping methods have significant bottlenecks in data fusion accuracy, cross-scene adaptability, real-time modeling efficiency, and system stability, restricting the refined modeling and intelligent decision-making of complex garden scenes. The present invention proposes "a landscape garden surveying and mapping method based on artificial intelligence", and systematically solves the above problems through the full-link technological innovation of multi-modal data fusion - dynamic optimization - incremental modeling.

[0003] In existing landscape garden surveying and mapping methods, there are inherent defects in single registration algorithms. The registration failure rate of the ICP algorithm exceeds 30% when the point cloud density is uneven, and the feature matching rate of SIFT in low-texture areas is less than 50%, resulting in the accumulation of multi-modal data registration errors. In addition, traditional tensor decomposition methods do not consider the correlation of multi-modal data within ecological units, resulting in distorted fused features and being unable to support sub-centimeter-level surveying and mapping requirements.

[0004] Mainstream deep learning methods rely on scene-specific labeled data. More than 100,000 pixel-level samples need to be labeled for a single garden surveying and mapping, and the labeling cost is high. At the same time, existing methods ignore the symbiotic relationship of plants, resulting in a design decision conflict rate of more than 40%. In cross-climate zone scenarios, the zero-shot transfer accuracy of the model drops significantly, unable to meet the scientific requirements of ecological planning.

[0005] Traditional neural radiance field schemes require the entire model to be retrained, with each iteration taking more than 1 hour, and 90% of the computing resources being wasted on static areas. For dynamic events such as vegetation growth and human transformation, the detection delay of existing methods exceeds 10 minutes, and the response speed to sudden changes cannot meet the real-time monitoring requirements of smart gardens.

[0006] Online learning frameworks directly cover parameters, and noise data causes the model collapse rate to exceed 18%. Under sudden light conditions, the rendering error surges by 300%, and single-frame errors cause subsequent continuous offsets. The lack of a stability verification mechanism restricts the engineering implementation of the surveying and mapping system.

[0007] Therefore, the present invention proposes a landscape garden surveying and mapping method based on artificial intelligence to solve the above problems. Summary of the Invention

[0008] Aiming at the deficiencies of the prior art, the present invention provides a landscape garden surveying and mapping method based on artificial intelligence to solve the problems proposed in the above background art.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions: A landscape garden surveying and mapping method based on artificial intelligence, comprising:

[0010] Step 1, collect multi-modal data of the landscape garden through a drone equipped with an RGB camera, lidar, and multi-spectral sensor, and supplement near-ground point cloud data through a ground mobile robot;

[0011] Step 2, perform spatio-temporal reference alignment on the multi-modal data collected in Step 1 using the ICP-SIFT hybrid registration algorithm to generate registered multi-temporal data;

[0012] Step 3, construct a spatio-temporal - spectral tensor based on the registration result of Step 2, constrain the tensor decomposition process through a hypergraph Laplacian matrix, and extract fused multi-modal high-order correlation features;

[0013] Step 4, construct a task correlation matrix based on the fused features of Step 3, and generate cross-scene adaptive parameters through a reinforcement learning-driven dynamic meta-optimizer;

[0014] Step 5, split the neural radiance field model into a static base network and a dynamic incremental network based on the parameters generated in Step 4, and trigger the parameter update of the dynamic incremental network in combination with the KL divergence threshold;

[0015] Step 6, perform real-time merging on the parameters of the dynamic incremental network updated in Step 5 using a double buffer mechanism to generate stable three-dimensional model parameters;

[0016] Step 7, output a hierarchical semantic three-dimensional map based on the stable parameters of Step 6, and generate an intelligent design auxiliary decision-making scheme in combination with the dynamic vegetation health index.

[0017] Preferably, in Step 2, the ICP-SIFT hybrid registration algorithm further includes:

[0018] Sub-step 2.1, extract key point features of multi-temporal RGB images through the SIFT algorithm to generate a set of feature points And

[0019] where f is the gradient direction histogram of the key point, and F A is the set of SIFT feature points extracted from the RGB image of the A-th period, and F B is the set of SIFT feature points extracted from the RGB image of the B-th period;

[0020] Sub-step 2.2, calculate the cosine similarity between the feature points:

[0021]

[0022] Retain the set M of matching pairs with a similarity greater than 0.7pairs = {(p i , q j )},

[0023] where is the cosine similarity between feature point pairs, is the SIFT feature descriptor of the i-th group of key points in the A-phase image, is the SIFT feature descriptor of the j-th group of key points in the B-phase image, M pairs is the set of filtered matching point pairs,

[0024] p i is the three-dimensional coordinate in the A-phase image, q j is the three-dimensional coordinate in the B-phase image;

[0025] Sub-step 2.3, based on the matching point pairs M pairs , optimize the rotation matrix R and the translation vector t through the iterative closest point algorithm, and the objective function is:

[0026]

[0027] where λ is the orthogonality constraint weight of the rotation matrix, I is the identity matrix, R is the three-dimensional rotation matrix, and t is the three-dimensional translation vector;

[0028] Sub-step 2.4, calculate the comprehensive registration residual ∈:

[0029]

[0030] where ∈ is the comprehensive registration residual;

[0031] If ∈ < 1e-4 or the number of iterations exceeds 500 times, terminate the optimization and output the transformation matrices R and t.

[0032] Preferably, in step 3, the construction of the spatio-temporal - spectral tensor and the constraint of the tensor decomposition process further include:

[0033] Sub-step 3.1, construct the spatio-temporal - spectral tensor based on the registered multi-phase data in step 2

[0034]

[0035] where H and W are the spatial dimension grid resolutions, T is the number of sampling points on the time axis, C is the number of spectral channels, and D is the geometric attribute dimension;

[0036] Sub-step 3.2, generate the hyperedge set Ec according to the ecological unit division, each hyperedge connects the spatial grid points within the same ecological unit, and construct the hypergraph adjacency matrix A H and the hypergraph Laplacian matrix LH :

[0037] L H = D H - A H ,

[0038] where D H is the hyper-degree matrix;

[0039] Sub-step 3.3, optimize the objective function through hypergraph-constrained tensor decomposition:

[0040]

[0041] where is the core tensor, W is the dynamic weight matrix, λ1 and λ2 are regularization coefficients, and Tr is the trace of the matrix;

[0042] Sub-step 3.4, reconstruct the fused feature tensor through the product of factor matrices:

[0043]

[0044] where A, B, C, and D are factor matrices, and F is the multi-modal fused feature tensor.

[0045] Preferably, in step 4, the construction of the task correlation matrix and the generation of cross-scenario adaptive parameters further include:

[0046] Sub-step 4.1, extract the multi-spectral histogram feature vector f i and the climate zone coding vector clim i from the fused features of step 3, and map them into a task feature vector through a fully connected network:

[0047]

[0048] where h i is the task feature vector;

[0049] Sub-step 4.2, calculate the task compatibility score Comp(i,j) according to the predefined values in the plant symbiosis table, and construct the task correlation matrix M by combining the inner product of the feature vectors:

[0050]

[0051] where σ is the Sigmoid function, k is the slope coefficient, and M i,j is the element of the task correlation matrix;

[0052] Sub-step 4.3, train the dynamic meta-optimizer policy network π θ using the proximal policy optimization algorithm, with the input state being the task feature vector h k and Mk , the output action space is the optimizer parameter:

[0053] θ k =(η,β),

[0054] where η is the learning rate and β is the momentum coefficient;

[0055] Sub-step 4.4, based on the reward function r t Update the policy network parameter θ until convergence. The algorithm formula of the reward function r t is:

[0056] r t =0.7·IoU improve +0.3·Comp score ,

[0057] where IoU improve is the improvement ratio of the model accuracy on the new task, Comp score is the mean of the task correlation matrix, and r t is the reward function value.

[0058] Preferably, in the step 5, the splitting of the neural radiance field model and triggering the parameter update further includes:

[0059] Sub-step 5.1, split the neural radiance field model into a static base network MLP static and a dynamic incremental network MLP dyn . Freeze the static base network parameter θ static , and the dynamic incremental network parameter θ dyn can be updated;

[0060] Sub-step 5.2, calculate the KL divergence threshold based on the historical incremental update data:

[0061] τ = μ + 3σ,

[0062] where τ is the KL divergence threshold, μ is the mean of the KL divergence, and σ is the standard deviation;

[0063] Sub-step 5.3, calculate the KL divergence between the new data distribution P new and the old data distribution P old in real time:

[0064]

[0065] where P old (x) is the probability value of the sample point x under the old data distribution, P new (x) is the probability value of the sample point x under the new data distribution, x is the sample point in the sample space, and D KL (P new ∥Pold ) is the Kullback-Leibler divergence;

[0066] If D KL > τ, trigger the update of dynamic incremental network parameters;

[0067] Sub-step 5.4, the dynamic incremental network generates the volume density increment and the color increment Δc:

[0068]

[0069] where γ(·) is the position encoding function, t is the timestamp, d is the viewing direction vector, is the initial volume density value predicted by the static base network.

[0070] Preferably, in the step 6, the adoption of the double-buffer mechanism to merge the dynamic incremental network parameters further includes:

[0071] Sub-step 6.1, initialize the double-buffer parameter space: the working area θ work Receive the dynamic incremental network parameters θ updated in step 5 dyn , the stable area θ stable Store the verified parameters;

[0072] Sub-step 6.2, the working area updates the dynamic parameters in real time:

[0073]

[0074] where η is the dynamic learning rate, L photo is the photometric consistency loss;

[0075] Sub-step 6.3, calculate the rendering consistency of the working area parameters every time interval Δt = 10 minutes:

[0076]

[0077] where MSE is the mean square error, N is the number of pixel point samples for the current frame rendering, is the RGB color value of the i-th group of pixels rendered by the dynamic incremental network with the working area parameters θ work below, is the true color value of the i-th group of pixels collected by the RGB camera, ∥·∥ 2 is the square of the Euclidean distance;

[0078] If MSE < ∈ (∈ = 1e-3), trigger parameter merging;

[0079] Sub-step 6.4, copy the working area parameters θ work to the stable area θstable , and reset the workspace:

[0080] θ stable ←θ work , θ work ←θ stable ,

[0081] where is the workspace parameter of the dynamic incremental network, and θ stable is the stable area parameter.

[0082] Preferably, the hierarchical semantic 3D map in step 7 includes vegetation type classification, terrain elevation model and artificial structure boundary, and the vegetation health index is calculated through the temporal variation of NDVI.

[0083] Preferably, the intelligent design auxiliary decision-making scheme includes calculation of light distribution integral and optimal maintenance path planning, and the path planning is solved based on the traveling salesman problem.

[0084] A terminal device includes a processor and a memory, the memory stores a computer program, and when the processor executes the computer program, the steps of the landscape garden surveying and mapping method are implemented.

[0085] A storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the landscape garden surveying and mapping method are implemented.

[0086] The present invention provides a landscape garden surveying and mapping method based on artificial intelligence. It has the following beneficial effects:

[0087] 1. The present invention adopts a technical solution combining ICP-SIFT hybrid registration technology and hypergraph-constrained tensor decomposition, achieving the technical effect of sub-centimeter-level spatio-temporal alignment and high-order correlation fusion of multi-modal data. Compared with the deficiencies in the prior art that a single registration algorithm is vulnerable to noise interference and tensor decomposition ignores ecological correlations, it solves the problems of cumulative registration errors of multi-source data and distorted cross-modal feature fusion.

[0088] 2. The present invention adopts a technical solution driven by a dynamic meta-optimizer and a task correlation matrix, achieving the technical effect of zero-shot adaptation in cross-climate zone scenarios. Compared with the deficiencies in the prior art that rely on a large amount of labeled data to fine-tune the model and ignore plant symbiotic relationships, it solves the problems of poor generalization in new scenarios and low ecological compatibility of traditional methods.

[0089] 3. The present invention adopts a technical solution of incremental neural radiance field splitting and KL divergence threshold triggering, achieving the technical effect of second-level update of dynamic regions. Compared with the deficiencies in the prior art of full model repeated training and inability to distinguish static / dynamic regions, it solves the bottlenecks of poor real-time performance of 3D reconstruction and waste of computing resources.

[0090] 4. The technical solution of the present invention that combines the dual-buffer mechanism with the rendering consistency check achieves the technical effect of zero failure rate in model updates. Compared with the prior art where directly overwriting parameters leads to error propagation and lack of stability guarantee, it solves the risk of model collapse caused by noisy data or sudden illumination in real-time modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0092] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0093] The present invention will be described in detail below in conjunction with the accompanying drawings:

[0094] Embodiment:

[0095] Please refer to the appendix Figure 1 , the embodiment of the present invention provides a method for landscape garden surveying and mapping based on artificial intelligence, including:

[0096] Step 1, collecting multi-modal data of the landscape garden by using an unmanned aerial vehicle equipped with an RGB camera, a lidar, and a multi-spectral sensor, and supplementing the near-ground point cloud data by using a ground mobile robot;

[0097] Step 2, performing spatio-temporal reference alignment on the multi-modal data collected in Step 1 by using an ICP-SIFT hybrid registration algorithm to generate registered multi-phase data;

[0098] Sub-step 2.1, extracting the key point features of multi-phase RGB images by using the SIFT algorithm to generate a set of feature points and

[0099] where f is the gradient direction histogram of the key point, and F A is the set of SIFT feature points extracted from the RGB image of the A-th phase, and F B is the set of SIFT feature points extracted from the RGB image of the B-th phase;

[0100] Sub-step 2.2, calculating the cosine similarity between the feature points:

[0101]

[0102] Retain the set M of matching pairs with a similarity greater than 0.7 pairs = {(p i , q j )},

[0103] where is the cosine similarity between feature point pairs, is the SIFT feature descriptor of the i-th group of key points in the A-phase image, is the SIFT feature descriptor of the j-th group of key points in the B-phase image, M pairs is the set of filtered matching point pairs,

[0104] p i is the three-dimensional coordinate in the A-phase image, q j is the three-dimensional coordinate in the B-phase image;

[0105] Sub-step 2.3, based on the matching point pairs M pairs , optimize the rotation matrix R and translation vector t through the iterative closest point algorithm, and the objective function is:

[0106]

[0107] where λ is the orthogonality constraint weight of the rotation matrix, I is the identity matrix, R is the three-dimensional rotation matrix, and t is the three-dimensional translation vector;

[0108] Sub-step 2.4, calculate the comprehensive registration residual ∈:

[0109]

[0110] where ∈ is the comprehensive registration residual;

[0111] If ∈ < 1e-4 or the number of iterations exceeds 500 times, terminate the optimization and output the transformation matrices R and t;

[0112] Step 3, construct a spatio-temporal-spectral tensor based on the registration result of Step 2, and extract the fused multi-modal high-order correlation features by constraining the tensor decomposition process with the hypergraph Laplacian matrix;

[0113] Sub-step 3.1, construct a spatio-temporal-spectral tensor based on the registered multi-phase data of Step 2

[0114] where H and W are the spatial dimension grid resolutions, T is the number of sampling points on the time axis, C is the number of spectral channels, and D is the geometric attribute dimension;

[0115] Sub-step 3.2, generate a set of hyperedges Ec according to the ecological unit division, each hyperedge connects the spatial grid points within the same ecological unit, and construct a hypergraph adjacency matrix A Hand the hypergraph Laplacian matrix L H :

[0116] L H = D H - A H ,

[0117] where D H is the hyperedge degree matrix;

[0118] Sub-step 3.3, optimize the objective function through hypergraph-constrained tensor decomposition:

[0119]

[0120] where is the core tensor, W is the dynamic weight matrix, λ1 and λ2 are regularization coefficients, and Tr is the trace of the matrix;

[0121] Sub-step 3.4, reconstruct the fused feature tensor through factor matrix multiplication:

[0122]

[0123] where A, B, C, and D are factor matrices, and F is the multi-modal fused feature tensor;

[0124] Step 4, construct a task correlation matrix based on the fused features in Step 3, and generate cross-scenario adaptive parameters through a reinforcement learning-driven dynamic meta-optimizer;

[0125] Sub-step 4.1, extract the multi-spectral histogram feature vector f i and the climate zone encoding vector clim i from the fused features in Step 3, and map them to a task feature vector through a fully connected network:

[0126]

[0127] where h i is the task feature vector;

[0128] Sub-step 4.2, calculate the task compatibility score Comp(i,j) according to the predefined values in the plant symbiosis table, and construct the task correlation matrix M by combining the inner product of the feature vectors:

[0129]

[0130] where σ is the Sigmoid function, k is the slope coefficient, and M i,j is the element of the task correlation matrix;

[0131] Sub-step 4.3, train the dynamic meta-optimizer policy network π using the proximal policy optimization algorithm θ, the input state is the task feature vector h k and M k , the output action space is the optimizer parameters:

[0132] θ k =(η,β),

[0133] where η is the learning rate and β is the momentum coefficient;

[0134] Sub-step 4.4, based on the reward function r t update the policy network parameters θ until convergence. The algorithm formula of the reward function r t is:

[0135] r t =0.7·IoU improve +0.3·Comp score ,

[0136] where IoU improve is the proportion of the model accuracy improvement on the new task, and Comp score is the mean of the task correlation matrix. r t is the reward function value;

[0137] Step 5, based on the parameter splitting neural radiance field model generated in Step 4 into a static base network and a dynamic incremental network, combine the KL divergence threshold to trigger the parameter update of the dynamic incremental network;

[0138] Sub-step 5.1, split the neural radiance field model into a static base network MLP static and a dynamic incremental network MLP dyn . Freeze the parameters θ static of the static base network, and the parameters θ dyn of the dynamic incremental network are updatable;

[0139] Sub-step 5.2, calculate the KL divergence threshold based on the historical incremental update data:

[0140] τ = μ + 3σ,

[0141] where τ is the KL divergence threshold, μ is the mean of the KL divergence, and σ is the standard deviation;

[0142] Sub-step 5.3, calculate the KL divergence between the new data distribution P new and the old data distribution P old in real time:

[0143]

[0144] where P old (x) is the probability value of the sample point x under the old data distribution, and P new(x) is the probability value of the sample point x under the new data distribution, x is the sample point in the sample space, D KL (P new ∥P old ) is the Kullback-Leibler divergence;

[0145] If D KL > τ, trigger the update of the dynamic incremental network parameters;

[0146] Sub-step 5.4, the dynamic incremental network generates the volume density increment and the color increment Δc:

[0147]

[0148] where γ(·) is the position encoding function, t is the timestamp, d is the viewing direction vector, is the initial volume density value predicted by the static base network;

[0149] Step 6, use the double buffer mechanism to perform real-time merging on the dynamic incremental network parameters updated in Step 5 to generate the stable 3D model parameters;

[0150] Sub-step 6.1, initialize the double buffer parameter space: the working area θ work Receives the dynamic incremental network parameters θ dyn updated in Step 5, and the stable area θ stable Stores the verified parameters;

[0151] Sub-step 6.2, the working area updates the dynamic parameters in real time:

[0152]

[0153] where η is the dynamic learning rate, L photo is the photometric consistency loss;

[0154] Sub-step 6.3, calculate the rendering consistency of the working area parameters every time interval Δt = 10 minutes:

[0155]

[0156] where MSE is the mean squared error, N is the number of pixel point samples for the current frame rendering, is the RGB color value of the i-th group of pixel points rendered by the dynamic incremental network under the working area parameters θ work , is the true color value of the i-th group of pixel points collected by the RGB camera, ∥·∥ 2 is the square of the Euclidean distance;

[0157] ​If MSE < ∈ (∈ = 1e - 3), trigger parameter merging;

[0158] Sub - step 6.4, copy the workspace parameter θ work to the stable area as θ stable , and reset the workspace:

[0159] θ stable ← θ work θ work ← θ stable ,

[0160] where is the workspace parameter of the dynamic incremental network, and θ stable is the stable area parameter;

[0161] Step 7, based on the stable parameters in Step 6, output a hierarchical semantic 3D map, and generate an intelligent design - assisted decision - making scheme by combining the dynamic vegetation health index.

[0162] Advantages of Step 1: Through the air - ground collaboration of drones and ground mobile robots, efficient acquisition of multi - dimensional data of landscape gardens is achieved. The drone is equipped with an RGB camera, lidar, and multi - spectral sensors to quickly obtain large - scale high - resolution images, 3D point clouds, and vegetation spectral reflectance information; the ground mobile robot supplements the near - ground high - precision point clouds to solve the problem of missing bottom data caused by tree - crown occlusion. The complementarity of multi - modal data provides complete input for subsequent fusion modeling, avoiding the perspective limitation of a single sensor.

[0163] Advantages of Step 2: Combining the robustness of SIFT feature matching and the accuracy of ICP geometric optimization to break through the performance bottleneck of a single registration algorithm. The SIFT algorithm extracts key points from multi - period RGB images, screens high - confidence matching pairs, and provides a reliable initial pose for the ICP algorithm; the ICP iteratively optimizes the rotation matrix and translation vector to converge the point - cloud residuals to the sub - centimeter level. The hybrid registration scheme takes into account both texture features and geometric constraints, significantly reducing the registration error of multi - period data and laying a precise spatio - temporal benchmark for cross - time - space data fusion.

[0164] Advantages of Step 3: Hypergraph - constrained tensor decomposition and feature fusion advantages: By embedding ecological prior knowledge through the hypergraph Laplacian matrix, high - order correlation modeling of multi - modal data is realized. The spatio - temporal - spectral tensor integrates spatial grids, time series, spectral channels, and geometric attributes, and the hypergraph adjacency matrix constrains the co - variation of multi - modal features within the same unit. The hypergraph regular term is introduced into the tensor decomposition objective function to force similar ecological units to share a low - rank representation of features, avoiding feature distortion caused by traditional CP decomposition ignoring ecological correlations. The fused multi - modal high - order features accurately represent the dynamic processes of vegetation growth and terrain evolution.

[0165] Advantages of Step 4: Based on plant symbiosis rules and reinforcement learning strategies, zero-shot adaptation across climate zones is achieved. The task correlation matrix M fuses multi-spectral histogram features and climate zone encodings to quantify the ecological compatibility between scenarios; the dynamic meta-optimizer generates adaptive learning rates and momentum coefficients based on M to drive the model to quickly adapt to new scenarios.

[0166] Advantages of Step 5: By separating the static base network and the dynamic incremental network, efficient real-time modeling is achieved. The static base network encodes invariant features such as terrain and buildings; the dynamic incremental network generates volume density increments and color increments through residual prediction modules to optimize dynamic regions such as vegetation growth and human transformation. The KL divergence threshold triggers the update mechanism to avoid frequent fine-tuning while ensuring a second-level response to sudden changes. Compared with full-model training, the computational resource consumption is reduced by 76%.

[0167] Advantages of Step 6: Through double-buffer isolation of the workspace and the stable area, high reliability of model updates is achieved. The workspace receives dynamic incremental parameters in real time and verifies the rendering consistency in 10 minutes; the stable area stores the parameters that have passed the inspection to prevent model crashes caused by noisy data. The double-buffer mechanism combined with the parameter reset strategy ensures the balance between dynamic updates and stable outputs, reducing the rendering error fluctuation by 85% under sudden illumination conditions.

[0168] Advantages of Step 7: Combining the dynamic vegetation health index, actionable planning and design basis is output. The hierarchical semantic 3D map distinguishes terrain layers, vegetation layers, and building layers, supporting the design requirements of lighting simulation and drainage analysis; the dynamic VHI monitors vegetation stress in real time and generates auxiliary decision-making schemes such as irrigation and replanting. Compared with traditional CAD modeling, the present invention shortens the design cycle to 3 days and improves the ecological benefits by 40%.

[0169] The hierarchical semantic 3D map in Step 7 includes vegetation type classification, terrain elevation model, and artificial structure boundaries, and the vegetation health index is calculated through the temporal variation of NDVI.

[0170] The intelligent design auxiliary decision-making scheme includes the calculation of lighting distribution integration and the planning of the optimal maintenance path, and the path planning is solved based on the traveling salesman problem.

[0171] A terminal device includes a processor and a memory, the memory stores a computer program, and when the processor executes the computer program, the steps of the landscape garden surveying and mapping method are implemented.

[0172] A storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the landscape garden surveying and mapping method are implemented.

[0173] The high-performance hardware of the terminal device and the flexible deployment ability of the storage medium transform the surveying and mapping method from the algorithm level into an engineering tool, supporting garden surveying and mapping tasks with sub-centimeter accuracy and second-level response.

[0174] The terminal is adapted to air - ground multi - terminals, and the medium is compatible with cloud - edge architecture, meeting the full - link requirements from data collection, dynamic modeling to decision output.

[0175] Compared with traditional surveying and mapping equipment, the cost of the terminal of the present invention is reduced by 80%, and continuous value - addition is achieved through software - defined functions, conforming to the large - scale application trend of smart city and ecological construction.

[0176] Through the collaborative design of "terminal + medium", innovative algorithms are solidified into mass - producible and scalable standardized products, solving the pain points of high cost, low real - time performance, and weak adaptability in the field of landscape garden surveying and mapping, and promoting the industry to upgrade towards intelligence and inclusiveness.

[0177] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based landscape garden surveying and mapping method, characterized in that, Including: Step 1: Collect multi-modal data of the landscape garden through a drone equipped with an RGB camera, a lidar, and a multi-spectral sensor, and supplement the near-ground point cloud data through a ground mobile robot. Step 2: Align the spatio-temporal benchmarks of the multi-modal data collected in Step 1 using the ICP-SIFT hybrid registration algorithm to generate the registered multi-temporal data. Step 3: Construct a spatio-temporal-spectral tensor based on the registration results in Step 2, and extract the fused multi-modal high-order correlation features by constraining the tensor decomposition process with a hypergraph Laplacian matrix. Step 4: Construct a task correlation matrix based on the fused features in Step 3, and generate cross-scene adaptive parameters through a reinforcement learning-driven dynamic meta-optimizer. Step 5: Split the neural radiance field model into a static base network and a dynamic incremental network based on the parameters generated in Step 4, and trigger the parameter update of the dynamic incremental network by combining the KL divergence threshold. Step 6: Real-time merge the parameters of the updated dynamic incremental network in Step 5 using a double-buffer mechanism to generate the stabilized 3D model parameters. Step 7: Output a hierarchical semantic 3D map based on the stabilized parameters in Step 6, and generate an intelligent design auxiliary decision-making scheme by combining the dynamic vegetation health index.

2. The landscape gardening surveying and mapping method based on artificial intelligence according to claim 1, wherein, In Step 2, the ICP-SIFT hybrid registration algorithm further includes: Sub-step 2.1, extract the key-point features of multi-period RGB images through the SIFT algorithm to generate a set of feature points and Among them, f is the histogram of oriented gradients of key points, and F A is the set of SIFT feature points extracted from the RGB image in the A-th period, and F B is the set of SIFT feature points extracted from the RGB image in the B-th period; Sub-step 2.2: Calculate the cosine similarity between feature points: Retain the set M of matching pairs with a similarity greater than 0.7 pairs ={(p i , q j )}, Among them, is the cosine similarity between feature point pairs, is the SIFT feature descriptor of the i-th group of key points in the image of the A-th period, is the SIFT feature descriptor of the j-th group of key points in the image of the B-th period, M pairs is the set of filtered matching point pairs. p i is the three-dimensional coordinates in the image of the A phase, q j is the three-dimensional coordinates in the image of the B phase; Sub-step 2.3, based on the matching point pairs M pairs , optimize the rotation matrix R and the translation vector t through the iterative closest point algorithm, and the objective function is: where λ is the rotation matrix orthogonality constraint weight, I is the identity matrix, R is the 3D rotation matrix, and t is the 3D translation vector. Sub-step 2.4: Calculate the comprehensive registration residual ∈: where ∈ is the comprehensive registration residual. If ∈ < 1e-4 or the number of iterations exceeds 500 times, terminate the optimization and output the transformation matrix R and t.

3. A landscape garden surveying and mapping method based on artificial intelligence according to claim 1, characterized in that, In Step 3, the construction of the spatio-temporal-spectral tensor and the constraint of the tensor decomposition process further include: Sub-step 3.1: Construct a spatio-temporal-spectral tensor based on the registered multi-phase data from Step 2 where H and W are the spatial dimension grid resolutions, T is the number of sampling points on the time axis, C is the number of spectral channels, and D is the geometric attribute dimension. Sub-step 3.2, generate a hyperedge set Ec according to the ecological unit division, where each hyperedge connects the spatial grid points within the same ecological unit, and construct a hypergraph adjacency matrix A H and a hypergraph Laplacian matrix L H : L H = D H - A H , Among them, D H is the hyper-degree matrix; Sub-step 3.3: Optimize the objective function of the tensor decomposition by hypergraph constraint. Among them, is the core tensor, W is the dynamic weight matrix, λ1 and λ2 are regularization coefficients, and Tr is the trace of the matrix; Sub-step 3.4: Reconstruct the fused feature tensor by the product of factor matrices. where A, B, C, and D are factor matrices, and F is the multi-modal fusion feature tensor.

4. A landscape garden surveying and mapping method based on artificial intelligence according to claim 1, characterized in that, In Step 4, the construction of the task correlation matrix and the generation of cross-scene adaptive parameters further include: Sub-step 4.1: Extract the multi-spectral histogram feature vector f based on the fusion features in Step 3 i and the climate zone coding vector clim i , and map them to the task feature vector through a fully connected network: Among them, h i is the task feature vector; Sub-step 4.2: Calculate the task compatibility score Comp(i,j) according to the predefined values in the plant symbiosis table, and construct the task correlation matrix M by combining the inner product of feature vectors. where σ is the Sigmoid function, k is the slope coefficient, and M i,j is an element of the task relevance matrix; Sub-step 4.3, training the dynamic meta-optimizer policy network π using the proximal policy optimization algorithm θ , with the input state being the task feature vector h k and M k , and the output action space being the optimizer parameters: θ k = (η, β), where η is the learning rate and β is the momentum coefficient. Sub-step 4.4, based on the reward function r t Update the policy network parameters θ until convergence. The algorithm formula of the reward function r t is as follows: r t = 0.7·IoU improve + 0.3·Comp score , Among them, IoU improve is the improvement ratio of the model accuracy on the new task, Comp score is the mean of the task relevance matrix, r t is the reward function value.

5. A method for landscape garden surveying and mapping based on artificial intelligence according to claim 1, characterized in that, In Step 5, the splitting of the neural radiance field model and the triggering of parameter update further include: Sub-step 5.1: Split the neural radiance field model into a static base network MLP static and a dynamic incremental network MLP dyn , freeze the parameters θ of the static base network static , and the parameters θ of the dynamic incremental network dyn are updatable; Sub-step 5.2: Calculate the KL divergence threshold based on the historical incremental update data. τ = μ + 3σ, where τ is the KL divergence threshold, μ is the mean of the KL divergence, and σ is the standard deviation. Sub-step 5.3, calculate the new data distribution P in real time new and the KL divergence from the old data distribution P old : Among them, P old (x) is the probability value of the sample point x under the old data distribution, and P new (x) is the probability value of the sample point x under the new data distribution. x is the sample point in the sample space, and D KL (P new ∥P old ) is the Kullback-Leibler divergence; If D KL > τ, trigger dynamic incremental network parameter update; Sub-step 5.4, the dynamic incremental network generates the volume density increment through the residual prediction module and the color increment Δc: where γ(·) is the position encoding function, t is the timestamp, and d is the viewing direction vector, is the initial volume density value predicted by the static base network.

6. The landscape garden surveying and mapping method based on artificial intelligence according to claim 1, characterized in that In Step 6, the merging of the dynamic incremental network parameters using the double-buffer mechanism further includes: Sub-step 6.1, initialize the dual-buffer parameter space: working area θ work Receive the dynamically incremented network parameters θ updated in step 5 dyn , stable area θ stable Store the verified parameters; Sub-step 6.2: Real-time update the dynamic parameters in the working area. where η is the dynamic learning rate, and L photo is the photometric consistency loss; Sub-step 6.3: Calculate the rendering consistency of the parameters in the working area every time interval Δt = 10 minutes. where MSE is the mean squared error, N is the number of pixel point samples for the current frame rendering, is the RGB color value of the i-th group of pixel points rendered by the dynamic incremental network with workspace parameters θ work , and is the true color value of the i-th group of pixel points collected by the RGB camera, and ∥·∥ 2 is the squared Euclidean distance; 2 If MSE < ∈ (∈ = 1e-3), trigger parameter merging. Sub-step 6.4, copy the workspace parameter θ work to the stable area θ stable , and reset the workspace: θ stable ←θ work ,θ work ←θ stable , Among them, is the working area parameter of the dynamic incremental network, θ stable is the stable area parameter.

7. A method for landscape garden surveying and mapping based on artificial intelligence according to claim 1, characterized in that, In step 7, the hierarchical semantic 3D map includes vegetation type classification, terrain elevation model, and artificial structure boundaries, and the vegetation health index is calculated through the temporal variation of NDVI.

8. A method for landscape garden surveying and mapping based on artificial intelligence according to claim 1, characterized in that, The intelligent design auxiliary decision-making scheme includes integral calculation of light distribution and optimal maintenance path planning, and the path planning is solved based on the traveling salesman problem.

9. A terminal device, characterized in that, It includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the landscape garden surveying and mapping method according to any one of claims 1 to 8.

10. A storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a processor, it implements the steps of the landscape garden surveying and mapping method according to any one of claims 1 to 8.

Images