Terrain decision point intelligent analysis method and system based on neural operator, and storage medium

Abstract

The invention discloses a terrain decision point intelligent analysis method and system based on a neural operator, and a storage medium. The method comprises the following steps: preprocessing multi-source static terrain data, unifying coordinates and resolution, and generating derived indexes such as a road distance field and a reachable cost function; inputting the standardized multi-channel raster data into a Fourier neural operator model, and outputting at least one continuous utility field of visibility, safety and accessibility; candidate points are generated based on morphological features and road topology, and forbidden areas, minimum spacing and category mutual exclusion hard constraints are applied; taking the compliance candidate point as an action space, and adopting a sequence combination optimization strategy with an action mask to select no more than K decision points; according to the method, automation, compliance and global optimal selection of terrain decision points are achieved, the point selection accuracy and efficiency are improved, and the method is suitable for scenes such as situation simulation, deployment deduction and path organization and is high in practicability and adaptability.

Description

Technical Field

[0001] This invention belongs to the field of geospatial information processing technology, specifically relating to an intelligent analysis method, system, and storage medium for terrain decision points based on neural operators. Background Technology

[0002] Traditional terrain feature recognition methods rely heavily on human experience or rule-based multi-index weighted methods (such as MCE). However, in application scenarios involving large-scale, complex terrain (significant elevation differences, dense road / water systems, and diverse land cover), these methods present the following prominent problems: 1. Multi-source data collaboration is difficult and sensitive to parameter settings. Differences in format and resolution of different data sources lead to poor data fusion results. 2. It is difficult to stably express potential utilities such as "visibility / exposure / accessibility" in continuous space, and it cannot accurately reflect the global trend and local details of the terrain; 3. The greedy Top-K strategy of "selecting the highest score" is prone to redundant coverage, constraint conflicts (such as violations of minimum spacing and no-entry zone regulations) and global suboptimal problems, making it difficult to meet the compliance requirements in practical applications; To this end, we propose an intelligent analysis method, system, and storage medium for terrain decision points based on neural operators. Summary of the Invention

[0003] The purpose of this invention is to provide a method, system, and storage medium for intelligent analysis of terrain decision points based on neural operators, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent analysis method for terrain decision points based on neural operators, comprising the following steps: (10) Acquire multi-source static terrain data and unify coordinates and resolution. Standardize the digital elevation model, slope, aspect, terrain undulation / roughness, road and water system vectors, and land cover data. At the same time, generate road distance field and view prior V0. (20) Input the standardized multi-channel raster data into the two-dimensional neural operator model to obtain at least one continuous utility field among visibility, security and accessibility; (30) Based on morphological features and road topology, discrete candidate points are generated, and hard constraints such as no-entry zones, minimum spacing and class mutual exclusion are applied to obtain a set of compliant candidate points; (40) Using the set of compliant candidate points as the discrete action space, construct a state that includes the set of selected points, the cover representation and the features of candidate points, and adopt a sequential combination optimization strategy with action masks to select no more than K decision points under hard constraints; (50) Output the coordinates of the decision point set, the sorting results, the coverage report and the compliance report, and generate interpretable visualization results of the utility field and input data.

[0005] Preferably, in step (10), the standardization process involves using projected coordinates and a pixel resolution of 5-30 meters to standardize the z-score and constructing an reachability cost function as a reachability reference feature. The formula for the reachability cost function is shown in formula (1): (1); Where C_move(x,y) represents the reachability cost function, and x and y represent the raster coordinates; α, β, and γ represent the weighting coefficients of slope cost, off-road penalty, and wading penalty, respectively. SlopeCost(S) represents the slope cost function, where S represents the slope value; OffRoadPenalty indicates off-road penalty; WaterCrossingPenalty refers to penalties for water-related offenses.

[0006] Preferably, in step (20), the input channels of the two-dimensional neural operator include at least eight channels from the digital elevation model, slope, sine of slope aspect, cosine of slope aspect, topographic relief index or topographic location index, road distance field, view prior V0 and land cover. The output channels include visibility, security, accessibility, and optional comprehensive utility. The two-dimensional neural operator adopts a Fourier neural operator structure. The Fourier neural operator has 3-6 layers, each layer contains 8-16 pairs of Fourier modes, the channel width is 64-96, and it adopts Gaussian error linear unit activation function, layer normalization, residual connection and frequency domain truncation processing.

[0007] Preferably, the training loss of the continuous utility field includes weighted L2 loss and gradient consistency regularization to preserve the terrain boundary structure, and the formula for the loss function is shown in formula (2): (2); Where 𝓛 represents the total loss value; This represents the summation operation; Represents the loss weighting coefficient for each utility field; It represents the predicted utility field, including visibility, security, and accessibility; Indicates the target utility field; Represents the square of the L2 norm; The weights represent the gradient consistency regularization coefficients; Represents the gradient operator; This indicates a predicted overall utility field; Indicates the overall utility field of the objective; This represents the L1 norm.

[0008] Preferably, in step (30), the candidate points are generated in the following way: Local extremum detection and nonmaximum suppression are performed in digital elevation models and curvature fields to extract peaks, saddles and ridge inflection points; Select intersections and endpoints on the road map, and take samples at equal intervals every L meters along the road; After merging the two types of points, geometric conflicts are eliminated according to the no-entry zone mask, minimum spacing dmin, and category mutual exclusion rules, controlling the size of candidate points per tile to be 200-1500; Where L represents the equidistant sampling interval of the road, and dmin is the minimum spacing between candidate points.

[0009] Preferably, step (30) further includes: A k-nearest neighbor graph is constructed among candidate points, with the surface movement cost as the edge weight. Features are extracted for each candidate point, including sampled values ​​of visibility, security, and accessibility, the maximum mean and variance of the neighborhood of radius R, and topological geometric features including distance to the nearest road, road grade, topological node degree, and distance to the outer edge of the residential area. The neighborhood features are then fused and reduced to a fixed dimension through a lightweight graph attention network.

[0010] Preferably, in step (40), the state representation includes a binary indication of the selected set, a bitmap or sparse hash of the covered roads or areas, and a feature matrix of all candidate points; the action space is a set of candidate point indices, with a mask providing feasible subsets in real time. The sequential combination optimization strategy adopts a strategy optimization algorithm with action mask and a pointer-type decoding structure, in which a candidate point is selected only from the set of possible actions at each step; Action masks are used to shield actions that violate hard constraints such as minimum spacing, no-entry zones, maximum number of decision point selections, and class exclusivity. The optimization process terminates when K is reached or the marginal gain is below a threshold; where K represents the maximum number of decision point selections. The objective function formula for the sequential combination optimization strategy is shown in formula (3): (3); Where J represents the optimization objective value; w1, w2, w3, w4, and w5 represent road coverage gain, area coverage gain, safety improvement value, redundancy, and reachability cost weighting coefficient, respectively. ΔCover_road represents the road coverage gain; ΔCover_area represents the area coverage gain; ΔSafety represents the safety enhancement value; Redundancy indicates redundancy; AccessCost represents the cost of reaching access; The coverage gain is achieved through a bitmap Boolean OR operation on the service range of the candidate points; Redundancy is the percentage of overlap with the selected point service / view. The reachability cost is derived from the reachability cost grid, and a compressed bitmap structure can be used to reduce computational complexity.

[0011] Preferably, in step (50), the output includes: the coordinates, categories, comprehensive scores, lists and rankings of marginal coverage gains and reachability costs of decision points, and compliance reports of minimum spacing, restricted areas, capacity and mutual exclusion rules, as well as interpretable heatmaps based on sensitivity analysis of utility field or policy inputs.

[0012] A terrain decision point intelligent analysis system based on neural operators is provided to implement a terrain decision point intelligent analysis method based on neural operators. The system includes a processor and a memory, wherein the memory stores program instructions that run on the processor, and the program instructions execute the following modules: The data preprocessing module is used to realize the unification, standardization and generation of derived indicators of multi-source terrain data in step (10); The neural operator module is used to implement the continuous utility field modeling in step (20); The candidate point generation module is used to implement the candidate point generation and hard constraint screening in step (30) of claim 1; The sequential combination optimization module is used to implement the decision point selection under the constraints of step (40); The results output module is used to implement the results output and visualization of step (50).

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs steps (10)-(50) of a terrain decision point intelligent analysis method based on neural operators.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. More accurate representation of terrain features: Through multi-source data collaborative preprocessing and Fourier neural operator modeling, it can stably output the equivalent fields of visibility, security and accessibility in continuous space, taking into account both the global trend and local details of the terrain, and providing a scientific basis for point selection; 2. Stricter hard constraint satisfaction: By using hard constraint designs such as no-entry zone masks, minimum spacing control, and class mutual exclusion, combined with action mask technology, illegal options are blocked throughout the entire process from candidate point generation to point selection optimization, ensuring that the output results are fully compliant and avoiding the constraint conflict problems of traditional methods; 3. Enhanced global optimality of point selection: Employing a sequential combination optimization strategy with action masks, using comprehensive value as the objective function, it avoids redundant coverage and global suboptimal problems in greedy point selection, improves the coverage efficiency of roads and regions, and reduces redundancy; 4. Higher practicality and adaptability: The entire method process is highly automated, the candidate point generation and optimization process is computationally efficient, and the weight coefficients can be flexibly adjusted according to tactical needs, adapting to various application scenarios such as situation simulation and deployment simulation. Moreover, the output results include interpretable visualization, which is convenient for practical application and verification. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of the intelligent analysis method for terrain decision points based on neural operators according to the present invention. Figure 2 This is a schematic diagram of the structure of the Fourier neural operator (two-dimensional neural operator) of the present invention; Figure 3 This is a schematic diagram of the point selection generation and constraint filtering process of the present invention; Figure 4 This is a schematic diagram of the sequential combination optimization process of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figures 1-4 The intelligent analysis method for terrain decision points based on neural operators provided by this invention includes the following steps: (10) Multi-source data preprocessing: Acquire multi-source static terrain data, including digital elevation model, slope, aspect, terrain undulation / roughness, road and water system vectors, and land cover data; perform unified coordinate and resolution processing on the above data, adopting projected coordinates and 5-30 meter pixel resolution, and simultaneously perform z-score standardization to ensure consistent data format and numerical range that matches modeling requirements; after standardization, generate road distance field and view prior V0, and construct reachability cost function as reachability reference feature, the formula of which is shown in formula (1): (1); Where C_move(x,y) represents the reachability cost function, which is used to quantify the travel cost of any grid in the terrain, and x and y represent the grid coordinates, corresponding to the spatial location in the terrain; α, β, and γ represent the weighting coefficients for slope cost, off-road penalty, and wading penalty, respectively, with default values ​​of 0.5, 0.3, and 0.2, which can be adjusted according to actual scenario requirements. SlopeCost(S) represents the slope cost function, where S represents the slope value, used to quantify the impact of slope on traffic. OffRoadPenalty represents the off-road penalty, with penalty coefficients of 1.0, 1.3, and ∞ for open areas, woodlands, and built-up areas, respectively. WaterCrossingPenalty refers to water crossing penalties, which apply a penalty mechanism to the main channel of the river and the 30-50 meter buffer zone. All penalties are normalized to the [0,1] interval. (20) Modeling of continuous utility fields: The standardized multi-channel raster data is input into a two-dimensional neural operator model. The model learns the mapping relationship between terrain features and utility attributes to obtain at least one continuous utility field among visibility, security, and accessibility. The input channels of the two-dimensional neural operator include at least eight channels from the digital elevation model, slope, sine of slope aspect, cosine of slope aspect, topographic relief index or topographic location index, road distance field, view prior V0 and land cover, to ensure comprehensive capture of the core features of the terrain such as elevation, morphology and topology. The output channels include visibility (quantifying the observation range advantage of terrain points), security (quantifying the concealment and protection capabilities of terrain points), accessibility (quantifying the ease of passage of terrain points), and optional comprehensive utility field (a comprehensive evaluation index that integrates the above three types of utility). The two-dimensional neural operator adopts a Fourier neural operator structure with 3-6 layers, each containing 8-16 pairs of Fourier modes and a channel width of 64-96. It also employs Gaussian error linear unit activation function, layer normalization, residual connection and frequency domain truncation to balance the model's fitting ability and computational efficiency while maintaining the integrity of the terrain boundary structure. (30) Candidate point generation and constraint screening: Based on morphological features and road topology, discrete candidate points are generated: local extremum detection and non-maximum suppression are performed in digital elevation model and curvature field to extract peak, saddle and ridge inflection points and capture key morphological nodes of the terrain. Select intersections and endpoints on the road map, and perform equidistant sampling every L meters along the road (L is the equidistant sampling interval of the road, with a default value of 100-200 meters) to cover the key nodes of the road network; After merging the two types of points, hard constraints are applied to eliminate geometric conflicts: filtering is performed based on prohibited area masks (excluding prohibited areas such as building buffers of 50 meters and water buffers), minimum spacing dmin (dmin is the minimum spacing between candidate points, with a default value of ≥100-150 meters), and category mutual exclusion rules (such as mutual exclusion between "ridge peaks" and "road intersections"). The size of candidate points per tile is controlled to be 200-1500, resulting in a set of compliant candidate points. Further feature enhancement is performed on candidate points: a k-nearest neighbor graph (k is the number of neighbors, ranging from 6 to 12) is constructed among candidate points, with the surface movement cost as the edge weight to establish the association between candidate points; features are extracted for each candidate point, including sampled values ​​of visibility, security, and accessibility, the mean, variance, and maximum value of the radius R neighborhood (R is the neighborhood radius, default value is 50-100 meters), as well as topological geometric features including distance to the nearest road, road level, topological node degree, and distance to the outer edge of the residential area; the neighborhood features are fused and reduced to a fixed dimension through a lightweight graph attention network to improve the discriminative power and modeling efficiency of the features; (40) Sequential combination optimization: Using the set of compliant candidate points as a discrete action space, a state representation is constructed. The state includes a binary indicator of the set of selected points (marking the selected candidate points), a bitmap or sparse hash of the covered roads or areas (recording the coverage of the selected points), and a feature matrix of all candidate points (the feature basis supporting the point selection decision). The action space is a set of candidate point indices, and a feasible subset is given in real time by a mask to shield actions that violate hard constraints. A sequential combination optimization strategy with action masks is adopted. The strategy uses a strategy optimization algorithm with action masks and a pointer-type decoding structure. In each step, only one candidate point is selected from the set of possible actions to ensure the compliance of the point selection process. The action mask is used to shield actions that violate hard constraints such as minimum spacing, no-entry zone, maximum number of decision points K and class mutual exclusion. The optimization process terminates when K is reached (K is the maximum number of decision points, which is set according to actual needs) or the marginal gain is lower than the threshold (the default threshold is the coverage gain of two consecutive steps <0.5%). The optimization objective function of the sequential combination optimization strategy is shown in formula (3): (3); Where J represents the optimization objective value, used to measure the overall value of the selected point; w1, w2, w3, w4, and w5 represent the road coverage gain, area coverage gain, security improvement value, redundancy, and reachability cost weighting coefficients, respectively. The default values ​​are 1.0, 0.3, 0.4, 0.6, and 0.2, which can be adjusted according to tactical needs. ΔCover_road represents the road coverage gain, which is the increase in road coverage due to the addition of a new selected point; ΔCover_area represents the area coverage gain, which is the increase in the coverage of the terrain area by the newly selected point; ΔSafety represents the safety improvement value, that is, the safety improvement brought about by adding a new selection point; Redundancy indicates redundancy; AccessCost represents the cost of reaching access; Coverage gain is achieved through a bitmap Boolean OR operation on the service range of candidate points; Redundancy is the percentage of overlap with the selected point service / view. The reachability cost comes from the reachability cost grid, and a compressed bitmap structure can be used to reduce computational complexity; (50) Output of results: Output the coordinates of the decision point set, the sorting results, the coverage report, and the compliance report; The decision point list includes key indicators such as coordinates, category, comprehensive score, marginal coverage gain and reachability cost; The coverage report quantifies the coverage effect of decision points on roads and areas; the compliance report includes the implementation status of minimum spacing, no-entry zones, capacity, and mutual exclusion rules; It also generates interpretable visualizations of the utility field and input data, including utility field heatmaps, coverage maps, and sensitivity analysis maps, which facilitate result verification and business interpretation.

[0018] The training loss for continuous utility fields includes weighted L2 loss and gradient consistency regularization to preserve the terrain boundary structure. The formula for the loss function is shown in Equation (2): (2); Wherein, 𝓛 represents the total loss value, which is used to measure the difference between the predicted utility field and the target utility field; This represents the summation operation; This represents the loss weighting coefficients for each utility field, with default values ​​of 1.0 for visibility, 0.7 for security, and 0.7 for reachability. It represents the predicted utility field, including visibility, security, and accessibility; The target utility field is represented by a fusion of multi-index weighted priors and a small amount of manual annotation. It represents the square of the L2 norm, used to quantify the overall difference between the predicted and target values; This represents the weight coefficient for gradient consistency regularization, with a default value of 0.1. This represents the gradient operator, used to calculate the gradient of the utility field; This indicates a predicted overall utility field; Indicates the overall utility field of the objective; This represents the L1 norm, used to quantify gradient differences and maintain the continuity of terrain boundaries.

[0019] This invention also provides an intelligent terrain decision point analysis system based on neural operators to implement the above method, including a processor and a memory, wherein the memory stores program instructions that run on the processor, and the program instructions execute the following modules: Data preprocessing module: used to realize the unification, standardization and generation of derived indicators (road distance field, view prior V0, reachability cost function) of multi-source terrain data in step (10); The neural operator module is used to implement the continuous utility field modeling in step (20) and output the utilities of visibility, security, and reachability through Fourier neural operators. Candidate point generation module: used to implement candidate point generation and hard constraint screening in step (30), including morphological and topological feature extraction, constraint removal, and feature enhancement; Sequential combination optimization module: used to implement the decision point selection under the constraints of step (40), and complete the global optimal point through the strategy optimization algorithm with action mask; Results output module: used to realize the results output and visualization of step (50), and generate decision point list, coverage report, compliance report and heat map.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to perform steps (10) to (50) of the above method.

[0021] This embodiment is a specific implementation and application of the present invention, as detailed below: This embodiment uses 250 sets of multi-source terrain data for model training and testing. The data covers a 10×10km synthetic area (250-meter pixels) and includes complex terrain scenes such as mountains, plains, water bodies, and built-up areas. Each set of data includes input channels such as digital elevation model, slope, aspect, terrain undulation index (TRI), terrain location index (TPI), road and water system vectors, land cover, and view prior V0. The specific parameter configurations are as follows: Data preprocessing: Projected coordinates use UTM with a pixel resolution of 10 meters. The z-score normalization parameters are based on training set statistics (mean and variance are the statistical values ​​of each channel in the training set, respectively). Missing pixels are filled with the median of a 5×5 neighborhood. The road distance field is generated using a grid distance transformation algorithm. The view prior V0 uses 16 isotropic azimuths and 3-layer pyramids (each layer ×2 downsampling) for ray accumulation. The forest transmission attenuation coefficient is 0.6, and slope occlusion threshold tanφ>0.7 is considered invisible. Fourier neural operator: 4 layers, each with a Fourier mode of (12,12), channel width 64→96, Gaussian error linear unit activation function, layer normalization and residual connections enabled by default, frequency domain truncation coefficient 0.9; loss function weight λ k =Visibility 1.0, Security 0.7, Accessibility 0.7, η=0.1; Target Utility Field It is constructed by fusing multi-indicator weighted priors (visibility ↑, accessibility ↑, risk of being far away ↓) with 30-50 manually labeled samples per tile; Candidate point generation: Morphological extraction parameters prominence≥10m, |κ|≥0.02m⁻¹, road equidistant sampling interval L=150m, minimum interval dmin=120m, candidate point size per tile controlled at 500-800; k-nearest neighbor graph k=8, neighborhood radius R=75m, lightweight graph attention network output feature dimension 64; Sequential combination optimization: Maskable-PPO hyperparameters γ=0.99, GAEλ=0.95, clip=0.2, learning rate 3e-4; pointer-type Transformer layers: 4, head: 8, hidden layers: 256, feedforward layers: 1024; optimization objective function weights: w1=1.0, w2=0.3, w3=0.4, w4=0.6, w5=0.2; termination threshold: two consecutive steps (ΔCover_road+0.5・ΔCover_area)<0.5%; Reachability cost function: α=0.5, β=0.3, γ=0.2; off-road penalty is 1.0 / 1.3 / ∞ for open ground / woodland / building respectively; wading penalty is applied to the main channel of the river and the 30-50 meter buffer zone. The specific parameters and weight settings are shown in Table 1 below: Table 1: .

[0022] The specific implementation process is as follows: Data preprocessing: The 250 sets of input data were unified with UTM coordinates and 10-meter resolution, z-score standardization was performed, and missing pixels were filled in; the road distance field and view prior V0 were generated, the reachability cost function Cmove(x,y) was calculated, and the standardized multi-channel raster data was obtained. Table 2 below shows the multi-source input data and derived indicator data: Table 2: Model training: Input multi-channel raster data into the Fourier neural operator model, use formula (2) as the loss function, train for 200 epochs, batch_size=128, optimize the model parameters, and make the predicted utility field U k With target utility field Minimize the differences, and ultimately output visibility, security, accessibility and overall utility. Candidate point generation and screening: Peaks, saddles, and ridge inflection points are extracted based on morphological features, and intersections, endpoints, and equidistant sampling points are extracted based on road topology. After merging, the points are screened using a no-entry zone mask, a minimum spacing of 120 meters, and category mutual exclusion rules to obtain a set of compliant candidate points. A k=8 nearest neighbor graph is constructed, and the features of the candidate points are extracted and reduced to 64 dimensions using a lightweight graph attention network. Sequential combination optimization: Initialize the state (selected point set is empty, the covered bitmap is all 0, and the candidate point feature matrix is ​​loaded). Through the Maskable-PPO algorithm with action mask, select decision points in the order of the objective function of formula (3), and block illegal actions at each step until K=20 decision points are selected or the marginal gain is lower than the threshold. Results output: Generate a list of decision points (including coordinates, category, comprehensive score, etc.), a coverage report (road coverage gain 23.4%, area coverage gain 18.7%), a compliance report (minimum spacing violation rate 0, number of violations in prohibited areas 0, capacity utilization rate 100%), and output utility field heat map and coverage visualization results; Table 3 below shows the output data: Table 3: Table 4 below shows a performance comparison of existing attention models and the neural operator model of this invention: Table 4: The test results of this embodiment show that, compared with existing attention models, the performance indicators of the method of the present invention are significantly improved: the mean loss (MSE) is 0.12, which is 3% lower than the 0.15 of the existing attention model; the accuracy is 90.9%, which is 5.9% higher than the 85.0% of the existing attention model; the F1 score is 0.893, which is 7.3% higher than the 0.82 of the existing attention model; and the recall rate is 88.8%, which is 8.8% higher than the 80.0% of the existing attention model.

[0023] In a 10×10km synthetic area test, compared with the greedy method, the proposed method improved road coverage by 14.6-23.4%, improved safety by 14.6-23.4%, reduced redundancy by 18.9-31.2%, and had zero hard constraint violations, thus verifying the effectiveness and engineering feasibility of the proposed method.

[0024] The beneficial effects of this invention are as follows: 1. More accurate representation of terrain features: Through multi-source data collaborative preprocessing and Fourier neural operator modeling, it can stably output the equivalent fields of visibility, security and accessibility in continuous space, taking into account both the global trend and local details of the terrain, and providing a scientific basis for point selection; 2. Stricter hard constraint satisfaction: By using hard constraint designs such as no-entry zone masks, minimum spacing control, and class mutual exclusion, combined with action mask technology, illegal options are blocked throughout the entire process from candidate point generation to point selection optimization, ensuring that the output results are fully compliant and avoiding the constraint conflict problems of traditional methods; 3. Enhanced global optimality of point selection: Employing a sequential combination optimization strategy with action masks, using comprehensive value as the objective function, it avoids redundant coverage and global suboptimal problems in greedy point selection, improves the coverage efficiency of roads and regions, and reduces redundancy; 4. Higher practicality and adaptability: The entire method process is highly automated, the candidate point generation and optimization process is computationally efficient, and the weight coefficients can be flexibly adjusted according to tactical needs, adapting to various application scenarios such as situation simulation and deployment simulation. Moreover, the output results include interpretable visualization, which is convenient for practical application and verification.

[0025] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A terrain decision point intelligent analysis method based on neural operators, characterized in that, Includes the following steps: (10) Acquire multi-source static terrain data and unify coordinates and resolution. Standardize the digital elevation model, slope, aspect, terrain undulation / roughness, road and water system vectors, and land cover data. At the same time, generate road distance field and view prior V0. (20) Input the standardized multi-channel raster data into the two-dimensional neural operator model to obtain at least one continuous utility field among visibility, security and accessibility; (30) Based on morphological features and road topology, discrete candidate points are generated, and hard constraints such as no-entry zones, minimum spacing and class mutual exclusion are applied to obtain a set of compliant candidate points; (40) Using the set of compliant candidate points as the discrete action space, construct a state that includes the set of selected points, the cover representation and the features of candidate points, and adopt a sequential combination optimization strategy with action masks to select no more than K decision points under hard constraints; (50) Output the coordinates of the decision point set, the sorting results, the coverage report and the compliance report, and generate interpretable visualization results of the utility field and input data.

2. The intelligent analysis method for terrain decision points based on neural operators according to claim 1, characterized in that: In step (10), the standardization process involves using projected coordinates and a pixel resolution of 5-30 meters to standardize the z-score and constructing an reachability cost function as a reference feature for reachability. The formula for the reachability cost function is shown in formula (1): （1）； Where C_move(x,y) represents the reachability cost function, and x and y represent the raster coordinates; α, β, and γ represent the weighting coefficients of slope cost, off-road penalty, and wading penalty, respectively. SlopeCost(S) represents the slope cost function, where S represents the slope value; OffRoadPenalty indicates off-road penalty; WaterCrossingPenalty refers to penalties for water-related offenses.

3. The intelligent analysis method for terrain decision points based on neural operators according to claim 2, characterized in that: In step (20), the input channels of the two-dimensional neural operator include at least eight channels from the digital elevation model, slope, sine of slope aspect, cosine of slope aspect, topographic relief index or topographic location index, road distance field, view prior V0 and land cover. The output channels include visibility, security, accessibility, and optional comprehensive utility. The two-dimensional neural operator adopts a Fourier neural operator structure. The Fourier neural operator has 3-6 layers, each layer contains 8-16 pairs of Fourier modes, the channel width is 64-96, and it adopts Gaussian error linear unit activation function, layer normalization, residual connection and frequency domain truncation processing.

4. The intelligent analysis method for terrain decision points based on neural operators according to claim 3, characterized in that: The training loss of the continuous utility field includes weighted L2 loss and gradient consistency regularization to preserve the terrain boundary structure. The formula for the loss function is shown in formula (2): （2）； Where 𝓛 represents the total loss value; This represents the summation operation; Represents the loss weighting coefficient for each utility field; It represents the predicted utility field, including visibility, security, and accessibility; Indicates the target utility field; Represents the square of the L2 norm; The weights represent the gradient consistency regularization coefficients; Represents the gradient operator; This indicates a predicted overall utility field; Indicates the overall utility field of the objective; This represents the L1 norm.

5. The intelligent analysis method for terrain decision points based on neural operators according to claim 4, characterized in that: In step (30), the candidate points are generated as follows: Local extremum detection and nonmaximum suppression are performed in digital elevation models and curvature fields to extract peaks, saddles and ridge inflection points; Select intersections and endpoints on the road map, and take samples at equal intervals every L meters along the road; After merging the two types of points, geometric conflicts are eliminated according to the no-entry zone mask, minimum spacing dmin, and category mutual exclusion rules, controlling the size of candidate points per tile to be 200-1500; Where L represents the equidistant sampling interval of the road, and dmin is the minimum spacing between candidate points.

6. The intelligent analysis method for terrain decision points based on neural operators according to claim 5, characterized in that: Step (30) also includes: A k-nearest neighbor graph is constructed among candidate points, with the surface movement cost as the edge weight. Features are extracted for each candidate point, including sampled values ​​of visibility, security, and accessibility, the maximum mean and variance of the neighborhood of radius R, and topological geometric features including distance to the nearest road, road grade, topological node degree, and distance to the outer edge of the residential area. The neighborhood features are then fused and reduced to a fixed dimension through a lightweight graph attention network.

7. The intelligent analysis method for terrain decision points based on neural operators according to claim 6, characterized in that: In step (40), the state representation includes a binary indication of the selected set, a bitmap or sparse hash of the covered roads or areas, and a feature matrix of all candidate points; the action space is a set of candidate point indices, with a mask providing feasible subsets in real time. The sequential combination optimization strategy adopts a strategy optimization algorithm with action mask and a pointer-type decoding structure, in which a candidate point is selected only from the set of possible actions at each step; Action masks are used to shield actions that violate hard constraints such as minimum spacing, no-entry zones, maximum number of decision point selections, and class exclusivity. The optimization process terminates when K is reached or the marginal gain is below a threshold; where K represents the maximum number of decision point selections. The objective function formula for the sequential combination optimization strategy is shown in formula (3): （3）； Where J represents the optimization objective value; w1, w2, w3, w4, and w5 represent road coverage gain, area coverage gain, security improvement value, redundancy, and reachability cost weighting coefficient, respectively. ΔCover_road represents the road coverage gain; ΔCover_area represents the area coverage gain; ΔSafety represents the safety enhancement value; Redundancy indicates redundancy; AccessCost represents the cost of reaching access; The coverage gain is achieved through a bitmap Boolean OR operation on the service range of the candidate points; Redundancy is the percentage of overlap with the selected point service / view. The reachability cost is derived from the reachability cost grid, and a compressed bitmap structure can be used to reduce computational complexity.

8. The intelligent analysis method for terrain decision points based on neural operators according to claim 7, characterized in that: In step (50), the output results include: the coordinates, categories, comprehensive scores, lists and rankings of marginal coverage gains and reachability costs of decision points, and compliance reports of minimum spacing, restricted areas, capacity and mutual exclusion rules, as well as interpretable heatmaps based on sensitivity analysis of utility field or policy inputs.

9. A terrain decision point intelligent analysis system based on neural operators, used to implement the terrain decision point intelligent analysis method based on neural operators as described in any one of claims 1-8, characterized in that, It includes a processor and a memory, wherein the memory stores program instructions that run on the processor, and the program instructions execute the following modules: The data preprocessing module is used to realize the unification, standardization and derivation index generation of multi-source terrain data in step (10) of claim 1; The neural operator module is used to implement the continuous utility field modeling of step (20) of claim 1; The candidate point generation module is used to implement the candidate point generation and hard constraint screening in step (30) of claim 1; The sequential combination optimization module is used to implement the decision point selection under the constraints of step (40) of claim 1; The result output module is used to implement the result output and visualization of step (50) of claim 1.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it is used to perform steps (10)-(50) of the intelligent analysis method for terrain decision points based on neural operators as described in any one of claims 1-8.

Images