A non-measured area electromagnetic clutter deduction method and system based on similar terrain areas

Abstract

The application discloses a non-measured area electromagnetic clutter deduction method and system based on similar terrain areas, and the method comprises the following steps: acquiring geographic information data of a target non-measured area, and matching one or more reference areas with similar terrain characteristics; extracting measured electromagnetic clutter data and corresponding environmental parameters of each reference area, and constructing a terrain-electromagnetic correlation feature mapping model by using a deep feature learning network; dynamically generating a correction factor suitable for the area according to the terrain-electromagnetic correlation feature mapping model; based on the correction factor, deducing the spatial clutter intensity distribution of the target non-measured area under a specified frequency band and polarization mode; performing statistical feature conformity test and spatial continuity optimization on the spatial clutter intensity distribution, and finally outputting the electromagnetic clutter deduction result of the non-measured area. By using the embodiment of the application, the accuracy and spatial continuity of the non-measured area electromagnetic clutter deduction can be improved, and the dependence on field measurement can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic environment sensing technology, and in particular to a method and system for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions. Background Technology

[0002] In the field of electromagnetic environment sensing and modeling, accurately acquiring the electromagnetic clutter distribution of a specific geographical area is crucial for applications such as wireless communication, radar detection, and spectrum management. Traditional methods mainly rely on field measurements, but are limited by geographical conditions, weather changes, and cost, making it difficult to conduct comprehensive real-time measurements of all areas. In recent years, clutter extrapolation methods based on existing measured data or single theoretical models have gradually developed, which can alleviate the dependence on field measurements to some extent. However, most existing methods struggle to effectively integrate terrain features and dynamic environmental parameters, lacking the ability to deeply explore and transfer the similar terrain-electromagnetic correlation mechanisms between different regions. Furthermore, most extrapolation models do not fully consider the dynamic coupling relationship between multipath propagation, scattering mechanisms, and complex geographical environments, resulting in insufficient spatial continuity and statistical reliability of clutter extrapolation results in non-measured areas, especially in areas with complex terrain. Therefore, there is an urgent need for a non-measured area electromagnetic clutter modeling method that can fully utilize known measured data, integrate multi-source environmental parameters, and achieve high reliability and adaptive extrapolation. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, in order to overcome the shortcomings of the prior art, improve the accuracy and spatial continuity of extrapolating electromagnetic clutter in non-measured areas, and reduce the dependence on on-site measurements.

[0004] One embodiment of this application provides a method for extrapolating electromagnetic clutter in unmeasured areas based on similar terrain regions. The method includes: Acquire geographic information data of the target non-measured area and match one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database; Measured electromagnetic clutter data and their corresponding environmental parameters for each reference area were extracted, and a terrain-electromagnetic correlation feature mapping model was constructed using a metric-based deep feature learning network. Based on the topographic-electromagnetic correlation feature mapping model, and combined with the real-time meteorological parameters and radiation source characteristics of the target non-measured area, a correction factor for the multipath propagation and scattering coefficient applicable to the area is dynamically generated. Based on the aforementioned correction factor, the spatial clutter intensity distribution of the target's non-measured region under a specified frequency band and polarization mode is deduced through a hybrid computation engine that couples parabolic equations and empirical scattering models. The statistical characteristic conformity test and spatial continuity optimization are performed on the spatial clutter intensity distribution, and the electromagnetic clutter inference results for the non-measured area are finally output.

[0005] Optionally, the step of acquiring geographic information data of the target non-measured area and matching one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database includes: Digital elevation model data and land cover type data of the target non-measured area are obtained through a geographic information system to generate a terrain feature dataset of the target area. Multi-scale feature extraction is performed on the terrain feature dataset of the target area, and terrain quantification indicators, including at least terrain roughness, mean slope and standard deviation of elevation, are calculated to generate terrain feature vectors. The similarity between the terrain feature vector and the terrain features of each region in the known electromagnetic clutter measurement database is calculated, and a preliminary list of similar regions is generated by using a metric method based on Euclidean distance. The preliminary list of similar regions is filtered based on a preset terrain similarity threshold, and the final set of matching reference regions is generated by comprehensively considering the area of ​​the region and the completeness of the measured data.

[0006] Optionally, the step of extracting measured electromagnetic clutter data and corresponding environmental parameters for each reference area, and constructing a terrain-electromagnetic correlation feature mapping model using a metric-based deep feature learning network, includes: The measured clutter intensity data, meteorological parameters at the corresponding acquisition time, and radar system parameters for each region in the reference region set are extracted from the known electromagnetic clutter measurement database to generate a multimodal dataset of the reference region. The reference region multimodal dataset is normalized and aligned to eliminate differences in the units and collection conditions of data from different regions, thereby generating a standardized reference dataset. A metric-based deep feature learning network is constructed. The terrain feature vector and electromagnetic clutter feature vector from the standardized reference dataset are respectively input into the two branches of the Siamese network to learn the mapping relationship between the two in the shared latent space and generate an initial feature mapping model. The initial feature mapping model is trained using a triplet loss function. The network parameters are optimized to make the electromagnetic features corresponding to similar terrains closer in the latent space, and finally a fully trained terrain-electromagnetic correlation feature mapping model is generated.

[0007] Optionally, the step of dynamically generating a correction factor for the multipath propagation and scattering coefficients of the target non-measured area based on the terrain-electromagnetic correlation feature mapping model, combined with real-time meteorological parameters and radiation source characteristics of the target non-measured area, includes: Collect real-time meteorological data of the non-measured area of ​​the target, including temperature, humidity and atmospheric pressure, and obtain the operating frequency band, polarization mode and transmission power parameters of the preset radiation source to generate a real-time parameter set of the target area; By fusing the real-time parameter set of the target area with the terrain feature vector, a comprehensive feature description vector of the target area is constructed, and a comprehensive feature vector of the target area is generated. The comprehensive feature vector of the target region is input into the well-trained terrain-electromagnetic correlation feature mapping model. Its representation in the latent space is calculated through forward propagation, and the electromagnetic features of the k nearest reference regions in the latent space are retrieved to generate a set of nearest neighbor electromagnetic features. Based on the nearest electromagnetic feature set, the multipath propagation attenuation factor and scattering cross section adjustment coefficient of the target region relative to the standard theoretical model are calculated by a weighted average algorithm, and finally a set of correction factors applicable to the region is dynamically generated.

[0008] Optionally, the step of extrapolating the spatial clutter intensity distribution of the target's non-measured region under a specified frequency band and polarization mode based on the correction factor, using a hybrid computational engine that couples parabolic equations and empirical scattering models, includes: The hybrid computing engine is initialized, the standard parabolic equation solver and empirical scattering model library are loaded, and the computing grid is set according to the digital elevation model data of the target area to obtain the initialized computing environment; Input the set of correction factors, the real-time parameter set of the target region, and the characteristic parameters of the radiation source into the hybrid computing engine, configure the boundary conditions of the parabolic equation and the input parameters of the empirical scattering model, and obtain the engine input configuration. The hybrid computing engine is executed first. The parabolic equation module is run to calculate the basic propagation field of electromagnetic waves under complex terrain. Then, the empirical scattering model is called and the correction factor is substituted to calculate the clutter scattering intensity of each grid point, generating a preliminary spatial clutter intensity distribution matrix. The preliminary spatial clutter intensity distribution matrix is ​​post-processed, including unit transformation and logarithmic scaling transformation, to generate a preliminary spatial clutter intensity distribution map of the target non-measured area under a specified frequency band and polarization mode.

[0009] Optionally, the step of performing statistical characteristic conformity testing and spatial continuity optimization on the spatial clutter intensity distribution, and finally outputting the electromagnetic clutter inference results for the non-measured area, includes: Statistical features are extracted from the preliminary spatial clutter intensity distribution map, and statistics including at least mean, variance, skewness and kurtosis are calculated to generate a distribution statistical feature vector. The statistical feature vector of the distribution is compared with the prior distribution of the statistical features of large sample clutter data in the known measured database. The hypothesis testing method is used to judge the statistical rationality of the inference results and generate a statistical conformity test report. According to the statistical conformity test report, if there are local statistical anomalies, the anisotropic diffusion filtering algorithm is used to smooth the preliminary spatial clutter intensity distribution map, optimize its spatial continuity and suppress unreasonable abrupt changes, and generate an optimized spatial clutter intensity distribution map. By integrating and optimizing the spatial clutter intensity distribution map, statistical compliance test report, and key parameters and correction factors of the deduction process, a final electromagnetic clutter deduction result report with a standardized format is generated.

[0010] Another embodiment of this application provides a non-measured region electromagnetic clutter extrapolation system based on a similar terrain area, the system comprising: The acquisition module is used to acquire geographic information data of the target non-measured area and match one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database; The extraction module is used to extract measured electromagnetic clutter data and their corresponding environmental parameters for each reference area, and to construct a terrain-electromagnetic correlation feature mapping model using a metric-based deep feature learning network. The generation module is used to dynamically generate correction factors for the multipath propagation and scattering coefficients of the target non-measured area based on the terrain-electromagnetic correlation feature mapping model and in combination with the real-time meteorological parameters and radiation source characteristics of the target non-measured area. The deduction module is used to deduce the spatial clutter intensity distribution of the target non-measured area under a specified frequency band and polarization mode based on the correction factor and through a hybrid calculation engine that couples the parabolic equation and the empirical scattering model. The output module is used to perform statistical characteristic conformity verification and spatial continuity optimization on the spatial clutter intensity distribution, and finally output the electromagnetic clutter inference results for the non-measured area.

[0011] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0012] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0013] Compared with existing technologies, the present invention provides a method for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, which can improve the accuracy and spatial continuity of extrapolation of electromagnetic clutter in non-measured areas and reduce the dependence on on-site measurements. Attached Figure Description

[0014] Figure 1A hardware structure block diagram of a computer terminal for a method of extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electromagnetic clutter inference system for non-measured areas based on similar terrain regions, provided in an embodiment of the present invention. Detailed Implementation

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The present invention first provides a method for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0017] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a method of extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, provided as an embodiment of the present invention. (See diagram for reference.) Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0018] See Figure 2 The present invention provides a method for extrapolating electromagnetic clutter in unmeasured areas based on similar terrain regions, which may include the following steps: S201, acquire geographic information data of the target non-measured area, and match one or more reference areas with similar terrain features from the known electromagnetic clutter measurement database; Specifically, digital elevation model data and land cover type data of the target non-measured area can be obtained through a geographic information system to generate a terrain feature dataset of the target area; The core of this step is to accurately collect core terrain data of the target area using a geographic information system, and to construct a basic dataset covering elevation and surface attributes, providing data support for subsequent terrain feature analysis. The specific implementation method is as follows: The selection of a Geographic Information System (GIS) should be adapted to the high-precision acquisition requirements of terrain data. Systems supporting multi-source geographic data integration should be prioritized, utilizing their built-in data interfaces to access authoritative geographic data services. For digital elevation model (DEM) data acquisition of non-measured target areas, core parameters must be clearly defined. The resolution should be set to 30 meters (balancing data accuracy and computational efficiency, suitable for mesoscale terrain analysis), and the elevation accuracy should be controlled within ±2 meters (meeting the requirements of electromagnetic clutter simulation for depicting terrain undulations). The data format should adopt the GeoTIFF standard format and include geographic coordinate information (WGS84 coordinate system). In the example, the target area is a mountainous edge zone. The DEM data obtained through the GIS covers an area from 118°20′ to 118°30′ east longitude and 30°15′ to 30°25′ north latitude. The data matrix dimension is 2000×2000 pixels, with each pixel corresponding to a 30m×30m area in the field, and the elevation values ​​ranging from 150 to 800 meters.

[0019] The acquisition of land cover type data must be completely consistent with the spatial scope of the digital elevation model (DEM) data. The classification standard adopts a first-level classification system (including six core types: forest, grassland, cultivated land, built-up land, water area, and bare land). The data accuracy requirement is a type identification accuracy rate of ≥90%. Data for the target area is retrieved through the land cover data thematic database of the Geographic Information System (GIS). The data format is also GeoTIFF, and each pixel location corresponds one-to-one with the pixel location in the DEM data. Each pixel is labeled with the corresponding land cover type code (1-6 correspond to the six types mentioned above). In the example, the land cover types of the target area are mainly forest (code 1) and grassland (code 2), accounting for 65% and 25% respectively, with scattered cultivated land (code 3) and built-up land (code 4), and no water area (code 5) or bare land (code 6).

[0020] The generation of the target area terrain feature dataset requires spatial alignment and integration of the data. First, geographic coordinate matching is used to ensure that the pixels of the digital elevation model data and the land cover type data are completely aligned (coordinate deviation ≤ 1 pixel). Then, a structured dataset is constructed, with each data record containing the geographic coordinates (x, y) of the pixel, the elevation value z, and the land cover type code c. To improve the efficiency of subsequent processing, the dataset is divided into blocks, with the overall 2000×2000 data divided into 20×20 sub-blocks (each sub-block is 100×100 pixels), and the sub-blocks overlap by 5 pixels (to avoid loss of edge features). Finally, a target area terrain feature dataset containing 400 sub-blocks is generated, with each sub-block accompanied by spatial extent and data statistics (maximum elevation, minimum elevation, and main land cover type).

[0021] Multi-scale feature extraction is performed on the terrain feature dataset of the target area, and terrain quantification indicators, including at least terrain roughness, mean slope and standard deviation of elevation, are calculated to generate terrain feature vectors. The core of this step is to mine the hierarchical features of the terrain through multi-scale analysis, quantify key terrain attributes, and construct feature vectors that can accurately represent terrain differences. The specific implementation method is as follows: Multi-scale feature extraction needs to cover three levels: micro, meso, and macro. Considering the influence range of electromagnetic clutter propagation, three feature scales are defined: small scale (100 meters, corresponding to a range of 3-4 pixels, depicting local terrain undulations), meso scale (1000 meters, corresponding to a range of 33-34 pixels, depicting regional terrain trends), and large scale (5000 meters, corresponding to a range of 166-167 pixels, depicting the overall terrain pattern). For each sub-block of the target region's terrain feature dataset, feature extraction is performed at each of the three scales to ensure the comprehensiveness and hierarchy of the features.

[0022] The calculation of terrain quantification indicators is based on digital elevation model (DEM) data. At each scale, three core indicators are calculated: terrain roughness, mean slope, and standard deviation of elevation. Terrain roughness is calculated using the elevation variation coefficient, with the formula R = σ_z / μ_z, where σ_z is the standard deviation of elevation values ​​within the scale range, and μ_z is the mean elevation. A larger R value indicates more rugged terrain (range [0,1], where 0 represents absolute flatness and 1 represents extreme ruggedness). The mean slope is calculated by first determining the slope of each pixel using DEM data (slope is defined as the angle between the pixel surface and the horizontal plane, ranging from [0°, 90°]), and then taking the arithmetic mean of the slopes of all pixels within the scale range. The standard deviation of elevation is calculated as the square root of the sum of the squares of the deviations of elevation values ​​from the mean within the scale range, representing the degree of elevation dispersion.

[0023] Example calculation process: Within a small-scale (100-meter) range of a sub-block in the target area, the elevation values ​​are 250, 252, 248, ..., 255 meters (a total of 16 pixels). The calculated μ_z = 251 meters, σ_z = 2.3 meters, therefore the terrain roughness R = 2.3 / 251 ≈ 0.009. The pixel slopes within this range are 8°, 10°, 7°, ..., 9°, with a mean slope of 8.5°; the standard deviation of elevation is 2.3 meters. Within a medium-scale (1000-meter) range, μ_z = 320 meters, σ_z = 35 meters, R = 35 / 320 ≈ 0.109; the mean slope is 12.3°; the standard deviation of elevation is 35 meters. Within a large scale (5000 meters), μ_z = 450 meters, σ_z = 120 meters, R = 120 / 450 ≈ 0.267; the mean slope is 18.7°; and the standard deviation of elevation is 120 meters.

[0024] The generation of terrain feature vectors requires integrating nine quantitative indicators across three scales (3 scales × 3 indicators). These indicators are concatenated in the following order: "small-scale roughness - small-scale mean slope - small-scale elevation standard deviation - mesoscale roughness - mesoscale mean slope - mesoscale elevation standard deviation - large-scale roughness - large-scale mean slope - large-scale elevation standard deviation," forming a 9-dimensional feature vector. In the example, the terrain feature vector of the aforementioned sub-block is [0.009, 8.5, 2.3, 0.109, 12.3, 35, 0.267, 18.7, 120]. The units of each element in the vector are dimensionless, degrees, and meters, respectively. Subsequent normalization processing is needed to eliminate dimensional differences. To ensure the representativeness of the feature vectors, average pooling is performed on the feature vectors of all sub-blocks in the target area to generate a global terrain feature vector for the entire non-measured area of ​​the target, serving as the core basis for subsequent similarity matching.

[0025] The similarity between the terrain feature vector and the terrain features of each region in the known electromagnetic clutter measurement database is calculated, and a preliminary list of similar regions is generated by using a metric method based on Euclidean distance. The core of this step is to quantify the terrain similarity between the target area and the areas in the measured database using Euclidean distance, and then filter out potential similar areas. The specific implementation method is as follows: Preprocessing of the electromagnetic clutter measurement database requires ensuring a uniform data format. Each measured region in the database has a pre-calculated 9-dimensional global terrain feature vector (consistent with the target region's feature vector dimension), along with auxiliary information such as region number, geographical location, area, and measured data volume. To eliminate the influence of different indicator dimensions on similarity calculations, the terrain feature vector of the target region and the terrain feature vectors of all regions in the database need to be normalized using the min-max normalization method. The formula is x_norm=(x-x_min) / (x_max-x_min), where x is the original indicator value, and x_min and x_max are the minimum and maximum values ​​of that indicator in the database, respectively. In the example, the "small-scale average slope" of the target region's feature vector is 8.5°. In the database, x_min=2° and x_max=30°. After normalization, x_norm=(8.5-2) / (30-2)=6.5 / 28≈0.232.

[0026] The core of Euclidean distance-based similarity calculation is to measure the spatial distance between two 9-dimensional vectors; the smaller the distance, the higher the terrain similarity. The Euclidean distance calculation formula is d = √[Σ(i=1 to 9)(x_i-y_i)²], where x_i is the i-th element of the normalized feature vector of the target region, y_i is the i-th element of the normalized feature vector of a measured region in the database, and the value range of d is [0,√9] = [0,3]. During the calculation, all N measured regions in the database (assuming N=1000) need to be traversed, and the Euclidean distance between the target region and each measured region is calculated, resulting in N distance values.

[0027] Example calculation process: The normalized feature vector of the target region is [0.03, 0.232, 0.012, 0.32, 0.35, 0.21, 0.65, 0.48, 0.38], and the normalized feature vector of region A in the database is [0.028, 0.225, 0.011, 0.31, 0.34, 0.20, 0.64, 0.47, 0.37]. Calculate the Euclidean distance d = √[(0.03 - 0.028)² + (0.232 - 0.225)² + ... + (0.3...]. 8-0.37)²]≈√[4e-6+4.9e-5+…+1e-4]≈0.035; The normalized eigenvector of region B is [0.15,0.42,0.08,0.52,0.61,0.45,0.82,0.75,0.62], and d≈0.48 is calculated; The normalized eigenvector of region C is [0.032,0.24,0.013,0.33,0.36,0.22,0.66,0.49,0.39], and d≈0.042 is calculated.

[0028] The preliminary list of similar regions is generated by sorting all measured regions in ascending order of Euclidean distance and selecting the top M regions with the smallest distance (M is set to 50 to balance the number of candidate regions with the efficiency of subsequent screening). In the example, after sorting by distance, region A (d=0.035), region C (d=0.042), region D (d=0.051)... are ranked in the top 50. The region numbers, distance values, and region information of these 50 regions are integrated to generate a preliminary list of similar regions. Each entry in the list contains four core pieces of information: "region number - Euclidean distance - region area - measured data volume", which provide a basis for subsequent screening.

[0029] The preliminary list of similar regions is filtered based on a preset terrain similarity threshold, and the final set of matching reference regions is generated by comprehensively considering the area of ​​the region and the completeness of the measured data.

[0030] The core of this step is to optimize the initial list of similar regions through multi-dimensional filtering conditions, ensuring that the final selected reference regions have the characteristics of terrain similarity, area suitability, and data reliability. The specific implementation method is as follows: The terrain similarity threshold needs to be set based on the statistical characteristics of a known electromagnetic clutter measurement database. By calculating the mean μ_d and standard deviation σ_d of the Euclidean distance between all regions in the database, the threshold is set to μ_d - 0.5σ_d (ensuring that the selected regions are highly similar to the target region). In the example, the mean Euclidean distance μ_d = 0.2 and the standard deviation σ_d = 0.15 between regions in the database, so the threshold is set to 0.2 - 0.5 × 0.15 = 0.125. The 50 regions in the initial similarity list are then filtered, retaining those with an Euclidean distance ≤ 0.125. In the example, the distance between the first 50 regions is ≤ 0.1, and all pass the similarity screening, proceeding to the subsequent comprehensive consideration stage.

[0031] The area considerations must ensure that the area difference between the reference area and the target area is within a reasonable range, avoiding significant differences in terrain patterns and electromagnetic clutter propagation characteristics due to excessive area differences. An area difference threshold of 30% is set, meaning the ratio of the reference area S_ref to the target area S_target must satisfy 0.7 ≤ S_ref / S_target ≤ 1.3. If the target area S_target, calculated using the geographic information system, is 100 square kilometers, then the reference area must be within the range of 70-130 square kilometers. In the example, in the initial list, area A has an area of ​​95 square kilometers (95 / 100=0.95, compliant), area C has an area of ​​105 square kilometers (105 / 100=1.05, compliant), area E has an area of ​​65 square kilometers (65 / 100=0.65, non-compliant), and area F has an area of ​​135 square kilometers (135 / 100=1.35, non-compliant). Areas that do not meet the area criteria are removed, leaving 38 areas for consideration in the completeness of the measured data.

[0032] The completeness of measured data requires setting two core indicators: data coverage and data time span. Data coverage is defined as the proportion of valid clutter data collection points to the total number of grid points within the measured area, with a threshold of ≥80% (ensuring data coverage of most of the area). Data time span is defined as the range of time for data collection, with a threshold of ≥6 months (ensuring data covers different meteorological conditions and improving model generalization ability). In the example, region A has a data coverage of 92% and a collection time span of 8 months (compliant); region C has a data coverage of 88% and a collection time span of 7 months (compliant); region G has a data coverage of 75% (incompatible); and region H has a data collection time span of 4 months (incompatible). After removing regions that do not meet the data completeness requirements, 25 regions remain.

[0033] The final set of matching reference regions requires re-sorting the remaining 25 regions and selecting the top K regions (K is set to 10 to balance the number of reference regions and computational efficiency) based on their Euclidean distance from smallest to largest. In the example, the 10 regions with the smallest distances (Region A, Region C, Region D, Region I…) are selected, and their detailed information (region number, geographical location, terrain feature vector, area, and measured data details) is integrated to generate the final set of matching reference regions. Each reference region in the set is accompanied by complete measured electromagnetic clutter data and corresponding environmental parameters, providing high-quality training data support for the subsequent construction of a terrain-electromagnetic correlation feature mapping model.

[0034] S202, extract the measured electromagnetic clutter data and their corresponding environmental parameters for each reference area, and use a metric-based deep feature learning network to construct a terrain-electromagnetic correlation feature mapping model; Specifically, measured clutter intensity data, meteorological parameters at the corresponding acquisition time, and radar system parameters for each region in the reference region set can be extracted from the known electromagnetic clutter measurement database to generate a multimodal dataset of the reference region. The core of this step is to accurately extract multi-dimensional correlation data from the reference region and construct a multi-modal data system covering electromagnetic properties, environmental conditions, and observation systems. This provides comprehensive and matching data support for subsequent model training. The specific implementation method is as follows: Extracting measured clutter intensity data requires focusing on core observation indicators. Spatial distributed clutter intensity data for each reference area should be retrieved from the database. The data acquisition frequency should be set to 10Hz (to ensure the capture of instantaneous fluctuation characteristics of clutter), and the data unit should be decibels and milliwatts (dBm). The spatial resolution of the acquisition should be consistent with the digital elevation model of the target area (30 meters). Each data point includes geographic coordinates (x, y), acquisition timestamp t, and clutter intensity value P. In the example, the measured clutter intensity data of reference area A covers the range of 118°15′-118°25′ east longitude and 30°10′-30°20′ north latitude. The acquisition time was from 08:00 to 18:00 on June 10, 2024, with a total of 36,000 data points. The clutter intensity values ​​range from -85dBm to -40dBm, with the clutter intensity concentrated in the forest area between -75dBm and -60dBm, and in the built-up area between -65dBm and -45dBm.

[0035] The extraction of meteorological parameters corresponding to the acquisition time must ensure time synchronization, meaning that each clutter intensity data point is matched with meteorological data at the same timestamp. Extracted indicators include temperature (°C), relative humidity (%), atmospheric pressure (hPa), and wind speed (m / s). The data comes from the built-in meteorological observation sub-database, and the acquisition interval is consistent with the clutter data (10Hz). In the example, the clutter data for reference area A at 10:00:05 on June 10, 2024, is matched with meteorological parameters of 28°C, 65% relative humidity, 1012hPa atmospheric pressure, and 3.2m / s, ensuring that the impact of environmental conditions on clutter propagation can be accurately correlated.

[0036] Extracting radar system parameters requires identifying the core characteristics of the observation equipment. As a supplement to the clutter data observation background, the extracted parameters include the radar operating frequency band (GHz), polarization (horizontal polarization H / vertical polarization V / circular polarization C), transmit power (kW), antenna gain (dBi), and beamwidth (°). These parameters are either fixed values ​​or finely adjusted according to the observation task and are stored in the radar configuration sub-library of the database. In the example, the radar parameters for reference area A are: operating frequency band 3 GHz (S-band), vertical polarization V, transmit power 20 kW, antenna gain 35 dBi, and beamwidth 1.5°. This combination of parameters determines the propagation characteristics and scattering response of electromagnetic waves and must be forcibly bound to the clutter data.

[0037] The generation of the multimodal dataset for the reference area requires the integration and structured organization of the data. A dual spatial-temporal indexing mechanism is employed, with each data entry using "geographic coordinates (x, y) - timestamp t" as the core index, linked to clutter intensity P, meteorological parameters (temperature T, humidity RH, pressure P_atm, wind speed v_wind), and radar parameters (frequency band f, polarization pol, power P_tx, gain G, beamwidth θ), forming a 12-dimensional multimodal data record. To avoid data redundancy, duplicate data at the same spatial location and timestamp are deduplicated (the deduplication threshold is clutter intensity difference ≤ 0.5 dBm). Ultimately, each reference area generates an independent multimodal dataset, which includes basic regional information (area, terrain type) and data quality labels (percentage of valid data, missing data rate). In the example, the dataset for reference area A has a valid data rate of 98.5% and a missing data rate of 1.5% (mainly due to momentary malfunctions of meteorological sensors).

[0038] The reference region multimodal dataset is normalized and aligned to eliminate differences in the units and collection conditions of data from different regions, thereby generating a standardized reference dataset. The core of this step is to eliminate the heterogeneity of cross-regional data through data preprocessing, ensuring the consistency of data in terms of units, collection frequency, and spatial range, thus providing high-quality input for deep feature learning. The specific implementation method is as follows: Data normalization processing employs appropriate normalization methods for different data types to eliminate dimensional differences. For continuous numerical data (clutter intensity P, temperature T, relative humidity RH, atmospheric pressure P_atm, wind speed v_wind, transmit power P_tx, antenna gain G), the min-max normalization method is used, with the formula x_norm=(x-x_min) / (x_max-x_min), where x is the original data value, and x_min and x_max are the minimum and maximum values ​​of the index in all reference areas, respectively. After normalization, the data range is unified to [0,1]. In the example, the clutter intensity P has x_min=-90dBm and x_max=-30dBm. At a certain data point, P=-60dBm, and after normalization, x_norm=(-60-(-90)) / (-30-(-90))=30 / 60=0.5; the temperature T has x_min=-10℃ and x_max=40℃. At a certain data point, T=25℃, and after normalization, x_norm=(25-(-10)) / (40-(-10))=35 / 50=0.7. For categorical data (polarization mode pol), one-hot encoding is used to convert it into a numerical vector. Horizontal polarization H is encoded as [1,0,0], vertical polarization V is encoded as [0,1,0], and circular polarization C is encoded as [0,0,1]. For discrete numerical data (working frequency band f), it is mapped to numerical values ​​according to frequency band intervals. For example, S-band (2-4GHz) is mapped to 0.2, C-band (4-8GHz) is mapped to 0.4, and X-band (8-12GHz) is mapped to 0.6.

[0039] Data alignment processing includes two dimensions: temporal alignment and spatial alignment. Temporal alignment aims to unify the data acquisition frequency of each reference area. For areas with acquisition frequencies higher than 10Hz (e.g., 15Hz), a downsampling method is used to retain one data point every 100ms (i.e., 10Hz), with the downsampling strategy being to take the average value within 100ms. For areas with acquisition frequencies lower than 10Hz (e.g., 5Hz), a linear interpolation method is used to supplement data points to ensure that the time interval is uniformly 100ms. In the example, the clutter data acquisition frequency of reference area B is 5Hz, and the timestamps are 08:00:00 and 08:00:02. The clutter intensity value is obtained by linear interpolation at 08:00:01. The interpolation formula is x_interp=x_prev+(x_next-x_prev)×(t_interp-t_prev) / (t_next-t_prev), where x_prev and x_next are adjacent original data values, t_prev and t_next are the corresponding timestamps, and t_interp is the interpolation timestamp. Spatial alignment aims to unify the spatial resolution and coordinate system of each reference area, converting all area data into the WGS84 coordinate system with a unified spatial resolution of 30 meters. For areas with a resolution higher than 30 meters (such as 20 meters), the average pooling method is used to merge 2×2 pixels into one 30-meter pixel; for areas with a resolution lower than 30 meters (such as 50 meters), the bilinear interpolation method is used to improve the resolution to 30 meters, ensuring that the spatial grids of different areas are fully adapted.

[0040] The generation of a standardized reference dataset requires data integration and quality verification. Normalized and aligned data from each reference region are organized in a hierarchical structure of "region number - spatial coordinates - timestamp." Each data entry contains a normalized 12-dimensional feature vector (clutter intensity 4-dimensional: original normalized value + spatial gradient value + time rate of change + statistical mean; meteorological parameters 4-dimensional: temperature + humidity + pressure + wind speed; radar parameters 4-dimensional: frequency band mapping value + first 2 bits of polarization coding vector + transmit power + antenna gain). The quality verification process removes outliers after normalization (data exceeding the [0,1] range, ≤0.5%) and data with alignment failures (spatial coordinate matching error >1 pixel, ≤0.3%). The final standardized reference dataset contains all reference regions. The dataset includes a preprocessing log recording the normalization parameters, alignment methods, and data removal ratios for each region. In the example, the standardized dataset contains 10 reference regions and 5 million valid data entries, with data integrity ≥99.2%.

[0041] A metric-based deep feature learning network is constructed. The terrain feature vector and electromagnetic clutter feature vector from the standardized reference dataset are respectively input into the two branches of the Siamese network to learn the mapping relationship between the two in the shared latent space and generate an initial feature mapping model. The core of this step is to build a twin network architecture to achieve accurate mapping between terrain features and electromagnetic clutter features in the shared latent space, laying the network foundation for subsequent correlation model construction. The specific implementation method is as follows: The core architecture of the metric-based deep feature learning network is a Siamese network, which contains two identical feature extraction branches with shared parameters. These branches are used to extract terrain features and electromagnetic clutter features, respectively. Finally, spatial alignment of the two types of features is achieved through metric learning. Each feature extraction branch adopts a hybrid structure of "convolutional layer + fully connected layer". The input to the terrain feature branch is a 9-dimensional terrain feature vector (the multi-scale terrain quantization index generated earlier), and the input to the electromagnetic clutter feature branch is a 6-dimensional electromagnetic clutter feature vector (the original normalized value of clutter intensity, spatial gradient value, temporal rate of change, statistical mean, frequency band mapping value, and polarization coding first digit extracted from the standardized dataset).

[0042] The specific structure of each feature extraction branch is as follows: First, the input vector dimension is mapped to 64 dimensions through a fully connected layer, using ReLU (Modified Linear Unit, formula f(x)=max(0,x) as the activation function to enhance the non-linear expressive power of the network); then, two convolutional layers are connected, with a kernel size of 3×3, a stride of 1, and padding of 1. The first convolutional layer has 128 output channels, and the second convolutional layer has 256 output channels, both using ReLU as the activation function; next, a global average pooling layer compresses the convolutional features into a 256-dimensional vector; finally, two fully connected layers are used to reduce the dimension to 128 and 64 dimensions respectively, resulting in the final feature embedding vector (latent space dimension is 64). The network optimizer uses the Adam optimizer, with an initial learning rate of 0.0001, momentum parameters β1=0.9 and β2=0.999 (to improve training stability), and a weight decay coefficient of 1e-5 (to suppress overfitting).

[0043] The mapping relationship between terrain features and electromagnetic clutter features is learned through a shared latent space. Terrain feature vectors and their corresponding electromagnetic clutter feature vectors from a standardized reference dataset are paired and input into two branches of the Siamese network. After extracting features through shared parameters, the two branches project the 64-dimensional terrain feature embedding vector and the 64-dimensional electromagnetic clutter feature embedding vector into the same shared latent space. The network training objective is to minimize the distance between similar sample pairs (terrain-electromagnetic feature pairs from the same reference region) in the latent space and maximize the distance between dissimilar sample pairs (terrain-electromagnetic feature pairs from different reference regions). In the initial stage, the network parameters are randomly initialized (using the Xavier initialization method to ensure consistent output variance across layers) to generate an initial feature mapping model.

[0044] Example learning process: The terrain feature vector of reference region A is [0.03, 0.232, 0.012, 0.32, 0.35, 0.21, 0.65, 0.48, 0.38], and the corresponding electromagnetic clutter feature vector is [0.5, 0.3, 0.2, 0.4, 0.2, 0.1]. After inputting into the Siamese network, the terrain branch outputs a 64-dimensional latent space vector H1=[h1_1,h1_2,…,h1_64], and the electromagnetic branch outputs a 64-dimensional latent space vector E1. =[e1_1,e1_2,…,e1_64]; the terrain feature vector of reference region C is [0.032,0.24,0.013,0.33,0.36,0.22,0.66,0.49,0.39], and the corresponding electromagnetic clutter feature vector is [0.52,0.32,0.21,0.42,0.2,0.1]. The output latent space vectors are H2=[h2_1,…,h2_64] and E2=[e2_1,…,e2_64]. In the initial model, the Euclidean distance between H1 and E1 is 0.8, the Euclidean distance between H2 and E2 is 0.75, and the Euclidean distance between H1 and E2 is 0.9. It initially shows a trend that the distance between samples of the same class is smaller than that between samples of different classes, but the distance difference does not reach the ideal effect and needs to be optimized by loss function in the future.

[0045] The initial feature mapping model is trained using a triplet loss function. The network parameters are optimized to make the electromagnetic features corresponding to similar terrains closer in the latent space, and finally a fully trained terrain-electromagnetic correlation feature mapping model is generated.

[0046] The core of this step is to guide network optimization through a triplet loss function, strengthen the correlation and matching between terrain and electromagnetic features, and improve the feature mapping accuracy of the model. The specific implementation method is as follows: The definition of the triplet loss function and the construction of samples are the core of training. A triplet consists of an anchor sample (A), a positive sample (P), and a negative sample (N). The anchor sample selects the terrain-electromagnetic feature pair (H_A, E_A) of a certain reference area; the positive sample selects another terrain-electromagnetic feature pair (H_A', E_A') of the same reference area with high terrain similarity to the anchor sample. The terrain similarity is determined by the Euclidean distance. When the distance ≤ 0.05, it is considered highly similar; the negative sample selects the terrain-electromagnetic feature pair (H_N, E_N) of a different reference area with low terrain similarity to the anchor sample. When the Euclidean distance of the terrain similarity ≥ 0.2, it is considered lowly similar. The formula for the triplet loss function is L_triplet = max(d(A, P) - d(A, N) + margin, 0), where d(・, ・) is the Euclidean distance between two feature vectors in the latent space, and margin is the interval threshold (set to 0.2 to ensure a sufficient distance difference between similar and dissimilar samples). When d(A, P) + margin < d(A, N), the loss value is 0 and the network does not need to be updated; otherwise, the loss value is positive, and the network adjusts the parameters through backpropagation.

[0047] Parameter configuration and execution process of the training process: The training dataset consists of 1 million triplets randomly selected from the standardized reference dataset. The batch size is set to 128 (to balance training efficiency and gradient stability), the total number of iterations is set to 50,000 times, and a learning rate decay strategy is adopted. The learning rate decays to 0.5 of the original value every 10,000 iterations (to avoid late training oscillations). During the training process, the validation set loss is calculated every 1,000 iterations (the validation set contains 200,000 independent triplets). When the validation set loss fluctuates less than 1e-4 for 500 consecutive iterations, it is determined that the model has converged and training stops. The training uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the parameters of each layer of the network through the chain rule, and uses the Adam optimizer to update the parameters. The optimization goal is to minimize the value of the triplet loss function.

[0048] Example training process: In the initial iteration (1st time), the triplet loss value is 0.35, the latent space distance d(A,P) between anchor sample A (reference region A) and positive sample P is 0.8, and the distance d(A,N) between anchor sample A and negative sample N (reference region F) is 0.9, so the loss value is 0.8-0.9+0.2=0.1; after 10,000 iterations, the loss value drops to 0.08, d(A,P)=0.45, d(A,N)=0.72, so the loss value is 0.45-0.72+0.2=0.07; after 35,000 iterations, the loss value stabilizes at 0.02, and the validation set loss fluctuates at 0.00008 (less than 1e-4) over 500 consecutive iterations, indicating model convergence. At this point, the electromagnetic feature distances corresponding to similar terrains are significantly reduced in the latent space. The electromagnetic feature latent space distance between reference region A and reference region C (with similar terrain) decreases from the initial 0.85 to 0.32, while the electromagnetic feature distance between reference region A and reference region F (with different terrain) increases from 0.9 to 0.85, thus achieving the training objective of "similar terrains have similar electromagnetic features, while different terrains have different electromagnetic features".

[0049] Performance validation and output of the fully trained model: Model performance is evaluated using feature matching accuracy and confusion matrix. Feature matching accuracy is defined as the proportion of correctly matched terrain-electromagnetic feature pairs to the total number of samples. A threshold is set where a latent space distance ≤ 0.4 is considered a successful match. In the example, the test set contains 500,000 feature pairs, and the matching accuracy reaches 96.8%. The matching accuracy for single terrain areas such as woodland and grassland is ≥ 97.5%, and the matching accuracy for complex mixed terrain areas is ≥ 95%, meeting the model accuracy requirements. The final output is a fully trained terrain-electromagnetic correlation feature mapping model. The model file includes network structure configuration, optimized parameters for each layer, normalized parameters, and latent space mapping rules, which can be directly used for subsequent latent space representation calculation of the comprehensive feature vector of the target region.

[0050] S203, Based on the terrain-electromagnetic correlation feature mapping model, and combined with the real-time meteorological parameters and radiation source characteristics of the target non-measured area, a correction factor suitable for the multipath propagation and scattering coefficient of the area is dynamically generated. Specifically, it can collect real-time meteorological data of non-measured areas of the target, including temperature, humidity and atmospheric pressure, and obtain the operating frequency band, polarization mode and transmission power parameters of the preset radiation source to generate a real-time parameter set of the target area; The core of this step is to accurately collect dynamic environmental parameters and inherent characteristic parameters of the radiation source in the target area, and construct a real-time parameter system covering environmental and observational dimensions. This provides basic data support for subsequent comprehensive feature construction and correction factor generation. The specific implementation method is as follows: Real-time meteorological data acquisition must ensure both timeliness and spatial representativeness. A collaborative acquisition approach using a distributed meteorological sensor network and an authoritative meteorological data service interface is employed. The acquisition frequency is set to 1Hz (to ensure the capture of real-time fluctuations in meteorological parameters and adapt to the dynamic characteristics of electromagnetic clutter propagation). Acquired parameters include temperature, relative humidity, and atmospheric pressure, supplemented by wind speed and direction parameters (to assist in correcting multipath propagation path calculations). Temperature acquisition accuracy is ±0.1℃ (°C); relative humidity acquisition accuracy is ±1% (%); atmospheric pressure acquisition accuracy is ±0.1hPa (hPa); wind speed acquisition accuracy is ±0.1m / s (m / s); and wind direction accuracy is ±1° (°). In the example, the real-time meteorological data of the target non-measured area at a certain moment is as follows: temperature 26.3℃, relative humidity 62%, atmospheric pressure 1013.2hPa, wind speed 2.5m / s, wind direction 305°, and the data timestamp is 14:30:05 on July 15, 2024. The collection points cover four key locations in the target area (regional center, edge woodland, open grassland, and building edge). The global meteorological parameters of the area are obtained by arithmetic averaging to ensure the spatial representativeness of the data.

[0051] Obtaining the preset radiation source characteristic parameters requires clarifying the core indicators of the observation task. These parameters are derived from the radiation source configuration scheme. The core indicators include the operating frequency band, polarization, and transmit power, supplemented by antenna gain and beamwidth parameters (to complete the basic characteristics of the radiation source for electromagnetic wave propagation). The operating frequency band is measured in gigahertz (GHz), and the specific frequency range must be specified. In the example, the preset radiation source operating frequency band is 3.5 GHz (S-band, covering 3.1-4.9 GHz, suitable for medium-range electromagnetic clutter observation). Polarization includes three types: horizontal polarization (H), vertical polarization (V), and circular polarization (C). The example uses vertical polarization (V), which has relatively low propagation loss in complex terrain. The transmit power is measured in kilowatts (kW), and in the example, the transmit power is 25 kW. The transmit power determines the initial radiation intensity of the electromagnetic wave. The antenna gain is measured in decibels of omnidirectional radiated power (dBi), and in the example, it is 38 dBi, characterizing the antenna's ability to focus electromagnetic waves. The beamwidth is measured in degrees (°), and in the example, it is 1.2°, determining the spatial coverage range of the electromagnetic wave.

[0052] The generation of the real-time parameter set for the target area requires data association, integration, and structured organization. Using "timestamp-region identifier" as the core index, real-time meteorological parameters (temperature T, relative humidity RH, atmospheric pressure P_atm, wind speed v_wind, wind direction θ_wind) are linked with radiation source characteristic parameters (operating frequency band f, polarization mode pol, transmit power P_tx, antenna gain G, beamwidth θ_beam) to form a 10-dimensional parameter record. To ensure data quality, the collected data undergoes validity verification, removing outliers (parameters with temperatures exceeding -40℃ to 60℃, humidity exceeding 0% to 100%, and pressure exceeding 900hPa to 1100hPa). In the example, the temperature data at a certain collection point is 45.6℃ (within the valid range), humidity is 89% (valid), and pressure is 985hPa (valid), all passing the verification. The final generated real-time parameter set includes a data acquisition log, recording the type of acquisition equipment, acquisition time, and data validity ratio. In the example, the real-time parameter set has a valid data ratio of 99.8%, meeting the needs of subsequent calculations.

[0053] By fusing the real-time parameter set of the target area with the terrain feature vector, a comprehensive feature description vector of the target area is constructed, and a comprehensive feature vector of the target area is generated. The core of this step is to integrate the static terrain features of the target area with the dynamic environment and radiation source features to construct a comprehensive feature vector with complete dimensions and complementary information, providing comprehensive feature support for subsequent model input. The specific implementation method is as follows: Preprocessing before feature fusion must ensure that the data formats of the two types of features are consistent. The target area terrain feature vector is a 9-dimensional static vector generated earlier (3 scales × 3 terrain quantification indicators: small-scale roughness, small-scale slope mean, small-scale elevation standard deviation, medium-scale roughness, medium-scale slope mean, medium-scale elevation standard deviation, large-scale roughness, large-scale slope mean, large-scale elevation standard deviation). The real-time parameter set is a 10-dimensional dynamic parameter set (including categorical and continuous parameters). First, the different types of parameters in the real-time parameter set are standardized, and the processing method is consistent with the normalization rules of the reference area dataset: continuous parameters (temperature T, relative humidity RH, atmospheric pressure P_atm, wind speed v_wind, transmit power P_tx, antenna gain G) are normalized using min-max, with the formula x_norm=(x-x_min) / (x_max-x_min), where x_min and x_max are the extreme values ​​of the corresponding parameters in the reference area dataset; categorical parameters (polarization mode pol) are encoded using one-hot encoding; discrete parameters (working frequency band f) are mapped to values ​​in the range [0,1] according to the frequency band interval, and the wind direction θ_wind is converted into sine and cosine values ​​(θ_wind_sin=sin(θ_wind×π / 180), θ_wind_cos=cos(θ_wind×π / 180)) and included in the continuous parameter processing.

[0054] Example preprocessing: The temperature in the real-time parameter set of the target area is 26.3℃, and the temperature in the reference area dataset is x_min=-10℃, x_max=40℃. After normalization, x_norm=(26.3-(-10)) / (40-(-10))=36.3 / 50=0.726; the relative humidity is 62%, x_min=10%, x_max=95%, and after normalization, x_norm=(62-10) / (40 ... ) / (95-10)=52 / 85≈0.612; the working frequency band 3.5GHz (S-band) is mapped to 0.25; the vertical polarization (V) one-hot encoding is [0,1,0]; the wind direction is 305°, θ_wind_sin=sin(305°×π / 180)=sin(5.323rad)≈-0.819, θ_wind_cos=cos(5.323rad)≈0.574. The preprocessed real-time parameter set is converted into a 12-dimensional normalized feature vector (temperature 0.726, humidity 0.612, pressure normalized value 0.58, wind speed 0.32, wind direction sine -0.819, wind direction cosine 0.574, frequency band mapping value 0.25, polarization coding [0,1,0] first 2 bits 0 and 1, transmit power normalized value 0.625, antenna gain normalized value 0.76).

[0055] Feature fusion employs a feature concatenation method, directly concatenating a 9-dimensional normalized terrain feature vector with a 12-dimensional normalized real-time parameter feature vector in the order of "terrain features first, real-time parameter features second," to generate a 21-dimensional comprehensive feature vector for the target area. The core logic of this concatenation is that static terrain features determine the basic environment for electromagnetic clutter propagation, while dynamic real-time parameters determine the dynamic correction factors for clutter propagation; the two complement each other to form a complete feature description system. In the example, the normalized terrain feature vector for the target area is [0.03, 0.232, 0.012, 0.32, 0.35, 0.21, 0.65, 0.48, 0.38], and the preprocessed real-time parameter feature vector is [0.726, 0.612, 0.58, 0.32, -0.819, 0.574, 0.25, 0, 1, 0.625, 0.76, 0.35]. (With a beamwidth normalization value of 0.35 added), the resulting 21-dimensional composite feature vector after stitching is [0.03, 0.232, 0.012, 0.32, 0.35, 0.21, 0.65, 0.48, 0.38, 0.726, 0.612, 0.58, 0.32, -0.819, 0.574, 0.25, 0, 1, 0.625, 0.76, 0.35]. After fusion, the feature vector needs to be validated to ensure that all elements are within [0, 1] or a reasonable numerical range (e.g., wind direction sine / cosine within [-1, 1]). After successful validation, the final target area composite feature vector is generated and used as input to the terrain-electromagnetic correlation feature mapping model.

[0056] The comprehensive feature vector of the target region is input into the well-trained terrain-electromagnetic correlation feature mapping model. Its representation in the latent space is calculated through forward propagation, and the electromagnetic features of the k nearest reference regions in the latent space are retrieved to generate a set of nearest neighbor electromagnetic features. The core of this step is to map the comprehensive features to the latent space through a fully trained model, and use nearest neighbor retrieval to find the electromagnetic features of the most similar reference region, providing a reference for subsequent correction factor calculation. The specific implementation method is as follows: The forward propagation computation of latent space representation is a process of feature extraction and mapping of comprehensive features by the model. A fully trained terrain-electromagnetic correlation feature mapping model includes a feature preprocessing layer and a feature mapping layer. First, the 21-dimensional comprehensive feature vector of the target region is input into the feature preprocessing layer of the model. This layer maps the 21-dimensional features to 64 dimensions through a fully connected layer, with the activation function being ReLU (formula f(x)=max(0,x)), achieving the unification and nonlinear transformation of feature dimensions. Then, the 64-dimensional features are input into the feature mapping layer (i.e., the shared feature extraction branch of the Siamese network). Through processing by two convolutional layers, one global average pooling layer, and two fully connected layers, a 64-dimensional latent space representation vector is finally output. This vector is a condensed expression of the comprehensive features in the terrain-electromagnetic correlation latent space, directly reflecting the correlation between the comprehensive characteristics of terrain, environment, and radiation source and the electromagnetic clutter characteristics of the target region.

[0057] Example forward propagation process: The 21-dimensional comprehensive feature vector of the target region is input into the preprocessing layer. The weight matrix of the fully connected layer is a 64×21 matrix (optimized parameters). After matrix multiplication and ReLU activation, a 64-dimensional intermediate feature vector [h1,h2,…,h64] is output, where h1=0.32, h2=0.15,…,h64=0.28. This intermediate feature vector is input into the feature mapping layer. After processing by the first convolutional layer (3×3 convolutional kernel, 128 output channels), a 128-dimensional feature is generated. After processing by the second convolutional layer (3×3 convolutional kernel, 256 output channels), a 256-dimensional feature is generated. The global average pooling layer compresses it into a 256-dimensional vector, and then it is reduced to 128-dimensional and 64-dimensional successively by two fully connected layers. Finally, the latent space representation vector H_target=[0.42,0.35,0.21,0.56,0.18,…,0.33] (64-dimensional) is output.

[0058] The nearest neighbor retrieval in the latent space employs a K-nearest neighbor retrieval algorithm based on Euclidean distance. The core of this algorithm is to calculate the Euclidean distance between the latent space representation vector H_target of the target region and all vectors in the latent space representation set of electromagnetic features of the reference regions, and then select the k vectors with the smallest distances. The latent space representation set of electromagnetic features of the reference regions is obtained by forward propagation of the electromagnetic feature vectors of all reference regions using a fully trained model. It contains 64-dimensional latent space vectors of electromagnetic features from 10 reference regions, denoted as H_ref_1, H_ref_2, ..., H_ref_10. The Euclidean distance is calculated as d_i = √[Σ(j=1 to 64)(H_target_j - H_ref_i_j)²], where H_target_j is the j-th element of H_target, and H_ref_i_j is the j-th element of the latent space vector of the i-th reference region. A smaller d_i indicates a higher degree of matching between the electromagnetic features of the reference region and the comprehensive features of the target region.

[0059] The setting of the k value needs to balance the representativeness of the reference samples and the stability of the calculation. Through statistical analysis of the number and similarity distribution of reference regions, k=5 is set (selecting 5 most similar reference regions, which ensures sample diversity while avoiding interference from too many samples). Example retrieval process: Calculate the Euclidean distance between H_target and the electromagnetic feature latent space vectors of 10 reference regions, obtaining d_1=0.18, d_2=0.32, d_3=0.15, d_4=0.25, d_5=0.22, d_6=0.45, d_7=0.38, d_8=0.16, d_9=0.28, d_10=0.36; sorted by distance from smallest to largest, d_3=0.15 (region 3) is obtained. d_8=0.16 (region 8), d_1=0.18 (region 1), d_5=0.22 (region 5), d_4=0.25 (region 4); Select the electromagnetic feature latent space vectors of the first 5 regions to generate the nearest neighbor electromagnetic feature set {H_ref_3,H_ref_8,H_ref_1,H_ref_5,H_ref_4}, with each vector accompanied by a corresponding Euclidean distance value, providing a weight basis for subsequent weighted averaging.

[0060] Based on the nearest electromagnetic feature set, the multipath propagation attenuation factor and scattering cross section adjustment coefficient of the target region relative to the standard theoretical model are calculated by a weighted average algorithm, and finally a set of correction factors applicable to the region is dynamically generated.

[0061] The core of this step is to utilize the electromagnetic characteristics of the nearest reference region and achieve dynamic estimation of the correction factor through weighted averaging, ensuring that the correction factor is adapted to the specific environment and radiation source characteristics of the target region. The specific implementation method is as follows: The weights are determined using an inverse distance weighting strategy. The core logic is that the smaller the Euclidean distance between the nearest neighbor reference region and the target region, the greater the reference value of its electromagnetic characteristics for the target region, and the higher its corresponding weight. The weight calculation formula is w_i=1 / d_i² / [Σ(m=1 to k)(1 / d_m²)], where w_i is the weight of the i-th nearest neighbor reference region, d_i is the Euclidean distance between that region and the target region, k=5 is the number of nearest neighbors, and the sum of the weights Σw_i=1 to ensure the rationality of the weighted average. Example weight calculation: The d_i values ​​for the nearest neighbor regions 3, 8, 1, 5, and 4 are 0.15, 0.16, 0.18, 0.22, and 0.25, respectively. The 1 / d_i² values ​​are 1 / (0.15²) = 44.444, 1 / (0.16²) = 39.0625, 1 / (0.18²) = 30.864, 1 / (0.22²) = 20.661, and 1 / (0.25²) = 16, respectively. The denominator Σ(1 / d_m²) = 44.444 + 39.0625 + 30.864 + 20.661 + 16 = 151.0315; therefore, the weights of each region are w_3 = 44.444 / 151.0315 ≈ 0.294, w_8 = 39.0625 / 151.0315 ≈ 0.259, w_1 = 30.864 / 151.0315 ≈ 0.204, w_5 = 20.661 / 151.0315 ≈ 0.137, and w_4 = 16 / 151.0315 ≈ 0.106. The sum of the weights is 0.294 + 0.259 + 0.204 + 0.137 + 0.106 = 1, which meets the requirements.

[0062] The multipath propagation attenuation factor is calculated based on the multipath attenuation value of the standard theoretical model, using a weighted average of the measured multipath attenuation corrections from the nearest reference region. The standard theoretical model uses a ray tracing model to calculate the multipath propagation attenuation value A_std (unit: dB). The measured multipath attenuation correction ΔA_i = A_meas_i - A_std_i from the nearest reference region, where A_meas_i is the measured multipath attenuation value in the reference region, and A_std_i is the value calculated by the standard theoretical model for that region. The target region multipath propagation attenuation factor α = 1 + Σ(w_i × ΔA_i) / A_std, where α is the correction coefficient. When α > 1, it indicates that the actual multipath attenuation is greater than the theoretical value, requiring an increase in the attenuation correction; when α < 1, it indicates that the actual attenuation is less than the theoretical value, requiring a decrease in the attenuation correction. Example calculation: The standard theoretical model calculates A_std=15dB for the target region; the ΔA_3=1.2dB for neighboring region 3, ΔA_8=0.8dB for region 8, ΔA_1=1.0dB for region 1, ΔA_5=0.6dB for region 5, and ΔA_4=0.9dB for region 4; the weighted average ΔA_avg=0.294×1.2+0.259×0.8+0.204×1.0+0.137×0.6+0.106×0.9≈0.353+0.207+0.204+0.082+0.095≈0.941dB; then α=1+0.941 / 15≈1+0.063≈1.063, indicating that the actual multipath attenuation in the target region is 6.3% higher than the theoretical value, and the theoretical multipath attenuation value needs to be corrected according to this factor.

[0063] The calculation of the scattering cross section adjustment factor is based on the standard radar cross section (RCS) model, and is a weighted average using the measured scattering cross section correction factors of the nearest reference region. The target region scattering cross section σ_std (unit: m²) is calculated by the standard theoretical model. The measured scattering cross section correction factor β_i = σ_meas_i / σ_std_i of the nearest reference region, where σ_meas_i is the measured scattering cross section of the reference region, and σ_std_i is the value calculated by the standard theoretical model for that region. The target region scattering cross section adjustment factor β = Σ(w_i × β_i), where β > 1 indicates that the actual scattering cross section is greater than the theoretical value, and β < 1 indicates that the actual scattering cross section is less than the theoretical value. Example calculation: β_3=1.12 for neighboring region 3, β_8=1.08 for region 8, β_1=1.15 for region 1, β_5=1.05 for region 5, and β_4=1.10 for region 4; the weighted average β=0.294×1.12+0.259×1.08+0.204×1.15+0.137×1.05+0.106×1.10≈0.329+0.280+0.235+0.144+0.117≈1.105, indicating that the actual scattering cross section of the target region is 10.5% higher than the theoretical value, and the theoretical scattering cross section needs to be corrected according to this coefficient.

[0064] The generation of the correction factor set integrates the multipath propagation attenuation factor and the scattering cross section adjustment coefficient, and supplements the environmental adaptation correction coefficient γ (a fine-tuning coefficient based on real-time meteorological parameters, γ=1+0.002×(T-25)-0.001×(RH-60), in the example T=26.3℃, RH=62%, γ=1+0.002×1.3-0.001×2=1+0.0026-0.002=1.0006), finally generating the correction factor set {α=1.063, β=1.105, γ=1.0006}. Each correction factor comes with its calculation basis (nearest neighbor region weight, measured data source, theoretical model type), ensuring the traceability and rationality of the correction factors. This set is directly used for the subsequent electromagnetic clutter intensity extrapolation of the hybrid computing engine.

[0065] S204. Based on the aforementioned correction factor, the spatial clutter intensity distribution of the target non-measured region under a specified frequency band and polarization mode is deduced through a hybrid calculation engine that couples the parabolic equation and the empirical scattering model. Specifically, the hybrid computing engine can be initialized, the standard parabolic equation solver and the empirical scattering model library can be loaded, and the computing grid can be set according to the digital elevation model data of the target area to obtain the initialized computing environment; The core of this step is to build an electromagnetic clutter simulation framework adapted to the target area, complete the loading of core algorithms and model libraries, and partition the computational space to provide a stable basic environment for subsequent simulations. The specific implementation method is as follows: The initialization of the hybrid computing engine requires clearly defining its core components. The engine adopts a modular architecture, including a data input module, a core computing module, and a result output module. During initialization, interface adaptation and parameter presets for each module must be completed. Loading the standard parabolic equation solver requires selecting a solution algorithm suitable for electromagnetic wave propagation calculations in complex terrain. The Split-Step Fourier Method (SSFM) is adopted as the core solution algorithm, which can efficiently handle electromagnetic wave propagation problems in non-uniform media and complex terrain boundaries, with solution accuracy meeting engineering requirements. During loading, the solver's core parameters must be initialized, including the maximum propagation distance (set to 1.2 times the diagonal length of the target area to ensure complete coverage of the entire area and surrounding influence range), the propagation step size (set to 0.5 meters to balance computational accuracy and efficiency), and the frequency adaptation range (covering the preset radiation source operating frequency band of 0.1-10 GHz).

[0066] Loading the empirical scattering model library requires integrating scattering models suitable for different land cover types. The library includes scattering models for forest land, grassland, built-up land, and bare land, etc. Each model is optimized based on a large amount of measured data and can be adaptively called according to the land cover type. During the loading process, a model index table needs to be established to bind the land cover type code with the corresponding scattering model. For example, forest land (code 1) is bound to the vegetation bistation scattering model, and built-up land (code 4) is bound to the rough surface specular reflection model, ensuring that the model can be quickly matched according to the land cover type of the target area.

[0067] The computational grid is set based on the digital elevation model (DEM) data of the target area. The core principle is to discretize the three-dimensional terrain of the target area into a two-dimensional planar grid, achieving accurate spatial mapping. The spatial resolution of the grid is consistent with the DEM (30 meters) to ensure that terrain undulation features are not lost. The grid covers the entire target area and extends outwards by 5 grid cells (to avoid computational errors caused by boundary effects). In the example, the DEM coverage of the non-measured target area is 118°20′-118°30′ east longitude and 30°15′-30°25′ north latitude, corresponding to a horizontal distance of approximately 18 kilometers east-west and 17 kilometers north-south. The computational grid is set to 610×580 (after expansion), with a grid cell step size of Δx = 30 meters in the x-direction and Δy = 30 meters in the y-direction. The coordinates (i,j) of each grid cell correspond to the actual location (x0 + i × Δx, y0 + j × Δy), where (x0, y0) are the geographic coordinates of the grid's starting point. Simultaneously, the elevation data from the digital elevation model is assigned to the corresponding grid cells to generate a grid elevation matrix H(i,j), providing terrain height input for solving the parabolic equation. Finally, the initialized computing environment includes a pre-loaded parabolic equation solver, an empirical scattering model library, and a computational grid bound to terrain information, which can directly receive input parameters to initiate inference calculations.

[0068] Input the set of correction factors, the real-time parameter set of the target region, and the characteristic parameters of the radiation source into the hybrid computing engine, configure the boundary conditions of the parabolic equation and the input parameters of the empirical scattering model, and obtain the engine input configuration. The core of this step is to complete the input and adaptation configuration of all parameters required for the deduction, clarify the calculation boundary rules and the basis for model calculation, and ensure that the core calculation module can accurately call the parameters to carry out the deduction. The specific implementation method is as follows: The parameter input process requires establishing a unified parameter interaction format. The set of correction factors (multipath propagation attenuation factor α, scattering cross section adjustment coefficient β, environmental adaptation correction coefficient γ), the real-time parameter set of the target area (temperature T, relative humidity RH, atmospheric pressure P_atm, etc.), and radiation source characteristic parameters (operating frequency band f, polarization mode pol, emission power P_tx, etc.) are organized hierarchically according to "calculation module - parameter type" and transmitted to the hybrid computing engine through the data input module. Parameter validity verification is required during input. For example, correction factors must be within a reasonable range of [0.8, 1.5] (to avoid extreme values ​​causing calculation distortion), and the radiation source operating frequency band must be within the solver's adaptation range. In the example, the set of correction factors {α=1.063, β=1.105, γ=1.0006} all passed verification, and the radiation source operating frequency band of 3.5GHz is within the 0.1-10GHz range. After successful verification, the parameters are stored in the engine parameter cache.

[0069] The boundary condition configuration for the parabolic equation requires clearly defining the input boundary, output boundary, and terrain boundary rules for electromagnetic wave propagation. The input boundary is the grid cell where the radiation source is located, configured as a Dirichlet boundary condition. The emission power and polarization of the radiation source are converted into the boundary electric field strength E0, with the formula E0=√(30×P_tx×G) / r, where r is the distance from the radiation source to the input boundary. In the example, P_tx=25kW, G=38dBi (converted to a linear value of 6309.57), and r=500 meters. The calculated E0=√(30×25000×6309.57) / 500≈√(4.732×10^9) / 500≈68800 / 500≈137.6V / m. The output boundary is the outer boundary of the computational grid, configured as an absorbing boundary condition, employing the Perfectly Matched Layer (PML) technique. The absorbing layer thickness is set to 3 grid cells, and the absorption coefficient increases linearly from 0 to 0.8, ensuring that electromagnetic waves propagate to the boundary without reflection and avoiding interference from boundary reflections with the internal field distribution calculation. The terrain boundary is the ground height of the grid cells. The grid elevation matrix H(i,j) is used as the lower boundary constraint of the parabolic equation, clearly defining the rules for electromagnetic wave reflection and transmission on the terrain surface. For example, when electromagnetic waves propagate to undulating terrain, the propagation direction is adjusted according to the terrain slope.

[0070] The input parameters of the empirical scattering model need to be adaptively adjusted according to the land cover type and real-time environmental parameters of the target area. Core parameters include surface roughness, dielectric constant, incident angle, frequency, polarization, and a set of correction factors. Surface roughness is assigned by the terrain roughness R calculated by the digital elevation model; the dielectric constant is determined based on the surface type and real-time humidity, for example, the dielectric constant for forest land is ε_r = 3.2 + j0.05 (at 62% humidity), and for built-up land, ε_r = 6.5 + j0.1; the incident angle is calculated from the location of the radiation source and the spatial orientation of the grid cell, using the formula θ = arctan(H(i,j) / d), where d is the horizontal distance from the grid cell to the radiation source; the frequency and polarization are directly extracted from the radiation source characteristic parameters. In this example, the frequency f = 3.5 GHz, and the polarization is vertical polarization, corresponding to the model input parameter pol = V. Simultaneously, the scattering cross-section adjustment coefficient β and the environmental adaptation correction coefficient γ are input into the model as correction terms for clutter scattering intensity calculation, ensuring that the model calculation results adapt to the actual environment of the target area. Finally, after all parameters are configured, an engine input configuration file is generated, which includes a parameter list, boundary condition rules, and a model parameter mapping table, providing complete parameter guidance for the execution of the hybrid computing engine.

[0071] The hybrid computing engine is executed first. The parabolic equation module is run to calculate the basic propagation field of electromagnetic waves under complex terrain. Then, the empirical scattering model is called and the correction factor is substituted to calculate the clutter scattering intensity of each grid point, generating a preliminary spatial clutter intensity distribution matrix. The core of this step is to complete the progressive deduction of "propagation field calculation - scattering intensity calculation" through a hybrid computing engine, and combine it with a correction factor to achieve accurate estimation of clutter intensity, thereby obtaining the spatial distribution of clutter intensity across the entire region. The specific implementation method is as follows: The calculation of the basic propagation field is dominated by the parabolic equation module. The core task is to solve for the propagation field distribution of electromagnetic waves under complex terrain and real-time weather conditions, obtaining the electric field intensity amplitude of each grid cell. The calculation process employs the split-step Fourier method, decomposing the electromagnetic wave propagation process into two stages: free space propagation and terrain / medium disturbance, which are then iteratively calculated sequentially. First, based on the electric field intensity E0 at the input boundary, the electric field intensity E1 of the electromagnetic wave propagating in free space to the first grid cell is calculated. Subsequently, combining the terrain height H(i,j) of this grid cell and the atmospheric medium parameters (the refractive index profile n(h) calculated from real-time temperature, humidity, and pressure), the electric field intensity is corrected to obtain E1'. This process is repeated, iteratively calculating grid by grid, ultimately obtaining the electric field intensity distribution matrix E(i,j) of the entire grid, i.e., the basic propagation field. In the example, the basic propagation field electric field strength E=85.2V / m for the central grid cell (i=305, j=290) of the target area, and E=42.6V / m for the edge grid cell (i=120, j=450) covered by forest, due to terrain shading and medium attenuation, accurately reflects the influence of complex terrain and environment on electromagnetic wave propagation.

[0072] The clutter scattering intensity is calculated based on the fundamental propagation field by calling an empirical scattering model and substituting a correction factor. First, the hybrid computing engine calls the corresponding scattering model from the empirical scattering model library according to the surface cover type encoding of each grid cell. Then, the electric field intensity E(i,j), surface roughness R, dielectric constant ε_r, incident angle θ, and other parameters of the fundamental propagation field of that grid cell are input into the model to calculate the uncorrected clutter scattering intensity S0(i,j), in watts per square meter (W / m²). Then, a correction factor is substituted to correct the intensity. The correction formula is S(i,j)=α×β×γ×S0(i,j), where α is the multipath propagation attenuation factor (correcting the attenuation deviation in the propagation process), β is the scattering cross section adjustment coefficient (correcting the theoretical deviation of the scattering intensity), and γ is the environmental adaptation correction coefficient (correcting the fine-tuning effect of the real-time environment). In the example, the grid cell (i=305, j=290) is grassland cover. The grassland scattering model is used to calculate S0=2.5×10^-8W / m². Substituting the correction factors α=1.063, β=1.105, and γ=1.0006, the calculation yields S=1.063×1.105×1.0006×2.5×10^-8≈1.063×1.105≈1.175, 1.175×1.0006≈1.176, and 1.176×2.5×10^-8≈2.94×10^-8W / m².

[0073] After calculating the clutter scattering intensity grid by grid, a preliminary spatial clutter intensity distribution matrix S is generated. The matrix dimension is consistent with the calculation grid (610×580), and each element S(i,j) corresponds to the clutter scattering intensity of that grid cell. Invalid grid cells (such as non-target area grids in the extended region) need to be marked in the distribution matrix using a value of -1.0×10^-20 W / m² for easy removal during subsequent post-processing. This matrix completely records the spatial distribution of clutter intensity in the non-measured target area under specified frequency bands and polarization modes, and is the core data foundation for subsequent post-processing and optimization.

[0074] The preliminary spatial clutter intensity distribution matrix is ​​post-processed, including unit transformation and logarithmic scaling transformation, to generate a preliminary spatial clutter intensity distribution map of the target non-measured area under a specified frequency band and polarization mode.

[0075] The core of this step is to transform the raw calculation results into a form that is easy to analyze and visualize through data post-processing, eliminating visualization distortion caused by excessive numerical differences, and generating an intuitive spatial clutter intensity distribution map. The specific implementation method is as follows: The core of the unit conversion is to convert the original clutter scattering intensity unit from watts per square meter (W / m²) to the commonly used decibels per milliwatt (dBm), which makes it easier to intuitively reflect the relative differences in clutter intensity. The conversion process consists of two steps: first, W / m² is converted to milliwatts per square meter (mW / m²), with the conversion relationship being 1W / m² = 1000mW / m²; then, milliwatts per square meter are converted to dBm, with the conversion formula being P_dBm = 10 × log10(S_mW / m²) + 10 × log10(A_e) - 10 × log10(4π), where A_e is the effective area of ​​the receiving antenna (calculated from the antenna gain G and operating frequency f in the radiation source characteristic parameters, with the formula A_e = G × λ² / (4π), λ is the electromagnetic wave wavelength, λ = c / f, and c is the speed of light 3 × 10^8 m / s). In the example, the S of a certain grid cell is 2.94 × 10⁻⁸ W / m², which is converted to S_mW / m² = 2.94 × 10⁻⁵ mW / m²; the radiating source antenna gain G = 38 dBi (linear value 6309.57), the operating frequency f = 3.5 GHz, λ = 3 × 10⁸ / (3.5 × 10⁹) ≈ 0.0857 m, and A_e = 6309.57 × (0.0857)² / (4 × 3.1416) ≈ 6309.57 × 0.00735 / 12.566≈6309.57×0.000585≈3.69m²; Substituting into the dBm conversion formula, P_dBm=10×log10(2.94×10^-5)+10×log10(3.69)-10×log10(12.566)≈10×(-4.53)+10×0.567-10×1.099≈-45.3+5.67-10.99≈-50.62dBm. For grid cells marked as invalid, they are still marked as invalid after unit conversion (e.g., -200dBm) to avoid interfering with valid data.

[0076] The core of logarithmic scaling is to compress the dynamic range of clutter intensity, solving the visualization ambiguity problem caused by excessively large differences in clutter intensity (potentially several orders of magnitude). A base-10 logarithmic transformation is used to further process the dBm value, with the transformation formula P_log=log10(P_dBm-P_min+1), where P_min is the minimum effective clutter intensity across the entire region (to avoid negative values ​​in the logarithmic transformation). In the example, the effective clutter intensity dBm value ranges from -85dBm to -40dBm, P_min=-85dBm, and for a certain grid cell P_dBm=-50.62dBm, after transformation P_log=log10(-50.62-(-85)+1)=log10(35.38)≈1.55. The transformed data range is compressed to [0, log10(85-40+1)] = [0, log10(46)] ≈ [0, 1.66], which facilitates subsequent visualization using color mapping.

[0077] The preliminary spatial clutter intensity distribution map is generated based on transformed clutter intensity data, using spatial rendering technology from a geographic information system (GIS). First, the transformed clutter intensity matrix is ​​bound to the geographic coordinates of the computational grid to ensure each data point accurately corresponds to its actual location. Then, a pseudo-color mapping scheme is used to map the compressed P_log values ​​to different colors; for example, P_log∈[0,0.5] is mapped to dark blue (low clutter intensity), P_log∈[0.5,1.0] to green (medium-low clutter intensity), P_log∈[1.0,1.5] to yellow (medium-high clutter intensity), and P_log∈[1.5,1.66] to red (high clutter intensity). Finally, the geographic contours of the target area (such as roads, rivers, and building boundaries) are overlaid to generate a preliminary spatial clutter intensity distribution map containing geographic information. The resolution of the distribution map is consistent with the computational grid (30 meters), clearly showing the spatial distribution patterns of clutter intensity. For example, built-up areas appear as red high clutter zones, and woodlands appear as green medium-low clutter zones, consistent with actual clutter propagation and scattering characteristics.

[0078] S205, the statistical characteristic conformity test and spatial continuity optimization of the spatial clutter intensity distribution are performed, and the electromagnetic clutter inference results of the non-measured area are finally output.

[0079] Specifically, statistical features can be extracted from the preliminary spatial clutter intensity distribution map, and statistics including at least mean, variance, skewness and kurtosis can be calculated to generate a distribution statistical feature vector; The core of this step is to extract key statistical indicators from the overall distribution of spatial clutter intensity, quantify the central tendency, dispersion, and distribution pattern of clutter intensity, and provide quantifiable comparative evidence for subsequent compliance testing. The specific implementation method is as follows: Statistical feature extraction requires a dBm-level clutter intensity data matrix (excluding invalid grid cells) corresponding to the preliminary spatial clutter intensity distribution map. Core extraction metrics include mean, variance, skewness, and kurtosis. Percentiles (10%, 50%, and 90th percentiles) and spatial gradient mean are supplemented to form a multi-dimensional statistical feature system. The mean reflects the central tendency of clutter intensity, calculated as μ=Σ(i=1 to M)Σ(j=1 to N)P_dBm(i,j) / (M×N), where M×N is the total number of valid grid cells, and P_dBm(i,j) is the clutter intensity dBm value of the (i,j)th grid cell. Variance reflects the dispersion of clutter intensity, calculated as σ²=Σ(i=1 to M)Σ(j=1 to N)[P_dBm(i,j)-μ]² / (M×N-1). Sample variance is used for calculation (denominator is M×N-1) to improve accuracy in small sample scenarios.

[0080] Skewness describes the asymmetry of clutter intensity distribution, calculated as S = Σ(i=1 to M)Σ(j=1 to N)[P_dBm(i,j)-μ]³ / [(M×N-1)×σ³]. When S>0, the distribution is right-skewed (long tail on the high clutter intensity side); when S<0, it is left-skewed (long tail on the low clutter intensity side); and when S=0, it is symmetrical. Kurtosis describes the steepness of the distribution, calculated as K = Σ(i=1 to M)Σ(j=1 to N)[P_dBm(i,j)-μ]. 4 / [(M×N-1)×σ 4 -3, by subtracting 3, the kurtosis of the normal distribution is reduced to 0. When K>0, the distribution is steeper (peaked) than the normal distribution, and when K<0, it is flatter (flat). The spatial gradient mean reflects the degree of spatial variation of clutter intensity. The calculation formula is G=Σ(i=1 to M)Σ(j=1 to N)√[(ΔP_x(i,j))²+(ΔP_y(i,j))²] / (M×N), where ΔP_x(i,j)=P_dBm(i+1,j)-P_dBm(i,j) is the gradient in the x-direction, and ΔP_y(i,j)=P_dBm(i,j+1)-P_dBm(i,j) is the gradient in the y-direction.

[0081] Example extraction process: The effective grid cell count of the preliminary spatial clutter intensity distribution map of the target area is 610×580-3200=340600 (after removing 3200 invalid extended grid cells). The calculated mean μ=-62.5dBm indicates that the overall clutter intensity in the area is concentrated around -62.5dBm; the variance σ²=36.2(dBm)², and the standard deviation σ≈6.02dBm reflects a moderate degree of clutter intensity dispersion; the skewness S=0.85, showing... The right-skewed distribution indicates the presence of a small number of high clutter intensity regions; the kurtosis K=0.32 shows a slightly peaked distribution, with the concentration of high clutter intensity regions slightly higher than in a normal distribution; the 10th quantile = -70.2dBm (10% of the grid clutter intensity is below this value), the 50th quantile (median) = -63.1dBm, and the 90th quantile = -55.8dBm; the spatial gradient mean G=0.85dBm / grid indicates that the spatial variation of clutter intensity is relatively gentle. The above eight statistical indicators are concatenated in the order of "mean-variance-skewness-kurtosis-10th quantile-50th quantile-90th quantile-spatial gradient mean" to generate an 8-dimensional distribution statistical feature vector [-62.5,36.2,0.85,0.32,-70.2,-63.1,-55.8,0.85].

[0082] The statistical feature vector of the distribution is compared with the prior distribution of the statistical features of large sample clutter data in the known measured database. The hypothesis testing method is used to judge the statistical rationality of the inference results and generate a statistical conformity test report. The core of this step is to verify the consistency between the statistical characteristics of clutter intensity obtained from the hypothesis testing and the prior distribution of the measured data, and to quantitatively judge the reliability of the hypothesis results. The specific implementation method is as follows: To construct the prior distribution of statistical features of large sample clutter data in a known measured database, it is necessary to select 1000 measured clutter data from regions similar to the target region in terms of terrain type, frequency band, and polarization. Extract the 8-dimensional statistical feature vector for each region, and use the kernel density estimation method to fit the prior probability distribution of each statistical index. At the same time, calculate the mean μ_prior, standard deviation σ_prior, and 95% confidence interval [μ_prior-1.96σ_prior, μ_prior+1.96σ_prior] of the prior distribution to form the prior distribution parameter set. In the example, the large sample of measured data close to the target area has a mean prior distribution of μ_prior=-63.2dBm, σ_prior=1.8dBm, and a 95% confidence interval of [-66.73,-59.67]dBm; and a skewness prior distribution of μ_prior=0.72, σ_prior=0.25, and a 95% confidence interval of [0.23,1.21].

[0083] The hypothesis testing method used is a combination of the Kolmogorov-Smirnov test and the t-test. The KS test is used to test whether the statistical characteristic distribution of the inferred results follows a prior distribution (overall distribution consistency test), and the t-test is used to test whether the mean of each individual statistical indicator is not significantly different from the mean of the prior distribution (individual indicator consistency test). The null hypothesis H0 of the KS test is: the distribution of the inferred statistical characteristics follows a prior distribution; the alternative hypothesis H1 is: it does not follow a prior distribution. The test statistic is D = max|F_emp(x) - F_prior(x)|, where F_emp(x) is the empirical distribution function of the inferred statistical characteristics, and F_prior(x) is the prior distribution function. The significance level is set to α = 0.05. If D ≤ D_α (the critical value corresponding to α = 0.05, determined by the sample size), then H0 is accepted, and the overall distribution is considered consistent.

[0084] The t-test is performed separately for each statistical indicator. The null hypothesis H0 is that there is no significant difference between the mean μ_emp and the prior mean μ_prior (μ_emp = μ_prior). The alternative hypothesis H1 is that there is a significant difference (μ_emp ≠ μ_prior). The test statistic t = (μ_emp - μ_prior) / (s / √n), where s is the sample standard deviation of the inferred indicator, n is the number of effective grid cells (considered as a large sample n > 30), and the significance level α = 0.05. If |t| ≤ t_α / 2(n-1) (the critical value of the t-distribution), then H0 is accepted, and the indicator is considered to be consistent with the prior distribution. Example test procedure: The extrapolated mean μ_emp = -62.5 dBm, the prior mean μ_prior = -63.2 dBm, s = 6.02 dBm, n = 340600, the calculated t = (-62.5 + 63.2) / (6.02 / √340600) ≈ 67.96, t_α / 2(340599) ≈ 1.96, |t| > 1.96, reject H0, indicating that the mean index is significant. Differences; Skewness μ_emp=0.85, μ_prior=0.72, s=0.15, calculated t=(0.85-0.72) / (0.15 / 583.6)≈505.8, |t|>1.96, reject H0; KS test statistic D=0.08, D_α=0.02 (critical value corresponding to sample size n=340600), D>D_α, reject H0, consider the overall distribution inconsistent.

[0085] The generation of a statistical conformity test report requires integrating the test results, including a test overview (test method, significance level, prior data source), individual indicator test results (t-value, critical value, and whether the test was passed for each indicator), overall distribution test results (KS statistic, critical value, and whether the test was passed), anomaly analysis (reasons for deviations in indicators that failed the test, such as a high mean possibly stemming from a deviation in the radiation source power assumption), and test conclusions (the statistical reasonableness level of the inference results, such as "average" or "poor"). In the example report, the test conclusion is: "Five indicators, including the mean and skewness, failed the t-test; the overall distribution failed the KS test; the statistical reasonableness level is average; there are local statistical anomalies, requiring spatial continuity optimization."

[0086] According to the statistical conformity test report, if there are local statistical anomalies, the anisotropic diffusion filtering algorithm is used to smooth the preliminary spatial clutter intensity distribution map, optimize its spatial continuity and suppress unreasonable abrupt changes, and generate an optimized spatial clutter intensity distribution map. The core of this step is to use an adaptive smoothing filtering algorithm to eliminate unreasonable abrupt changes in clutter intensity in statistically abnormal areas. While preserving the clutter variations caused by the actual terrain, it improves the continuity of spatial distribution. The specific implementation method is as follows: The core principle of the anisotropic diffusion filtering algorithm is to adaptively adjust the diffusion intensity based on the spatial gradient of clutter intensity. Strong diffusion (smoothing) is performed in regions where clutter intensity changes gradually (true continuous regions), while weak diffusion (preserving edge features and avoiding over-smoothing) is performed in regions where clutter intensity changes abruptly (potentially anomalous regions). This differs from the uniform diffusion of traditional Gaussian filtering, effectively balancing smoothing and feature preservation. The core iterative formula of the algorithm is P^(t+1)(i,j)=P^(t)(i,j)+λ×div[c(||∇P^(t)(i,j)||)×∇P^(t)(i,j)], where P^(t)(i,j) is the clutter intensity value in the t-th iteration, λ is the diffusion step size (controlling the smoothing degree in each iteration), ∇P^(t)(i,j) is the spatial gradient of clutter intensity, c(・) is the diffusion coefficient function, and div(・) is the divergence operator.

[0087] The diffusion coefficient function used is the Perona-Malik function, with the formula c(g)=1 / (1+(g / g0)²), where g=||∇P^(t)(i,j)|| is the gradient magnitude, and g0 is the gradient threshold (controlling the decay rate of the diffusion coefficient). The larger g0 is, the more uniform the diffusion; the smaller g0 is, the stronger the suppression of diffusion in abrupt regions. Parameter settings need to be combined with the spatial gradient characteristics of the preliminary distribution map. In the example, based on the spatial gradient mean G=0.85dBm / grid, g0=1.2dBm / grid (slightly larger than the gradient mean to ensure that only abrupt changes exceeding the normal gradient are suppressed), diffusion step size λ=0.15 (to avoid excessive smoothing due to an excessively large step size), and iteration count t_max=10 (gradually smoothing through multiple iterations to avoid distortion in a single iteration).

[0088] The filtering process is as follows: First, extract the dBm-level data matrix of the preliminary spatial clutter intensity distribution map and mark the statistically abnormal areas (grids around building sites with high mean values ​​according to the inspection report); then initialize the iteration count t=0 and use the preliminary data matrix as the initial iteration matrix P^(0); calculate P^(t+1)(i,j) grid by grid according to the iteration formula. During the calculation, the gradient ∇P^(t)(i,j) is calculated through the intensity difference between adjacent grids, the divergence div(・) is calculated through the spatial rate of change of the gradient, and the diffusion coefficient c(g) is adaptively adjusted according to the gradient magnitude; repeat the iteration until t=t_max to obtain the filtered clutter intensity data matrix. Example optimization results: The initial clutter intensity of the edge grid of a building site was -42dBm, while the surrounding grid was -58dBm, showing an unreasonable abrupt change of 16dBm. After filtering, the intensity of this grid became -48dBm, and the intensity of the surrounding grid became -55dBm, reducing the abrupt change to 7dBm, resulting in a smooth spatial transition. Meanwhile, the boundary grid of woodland and grassland (gradient 0.9dBm / grid, within the normal range) showed an intensity change of only 0.3dBm after filtering, preserving the edge features intact. The filtered data matrix was then logarithmically scaled and bound to geographic coordinates using the post-processing method in step one, generating an optimized spatial clutter intensity distribution map. The map showed a significantly improved match between high clutter areas and terrain / surface type, with no obvious unreasonable abrupt changes.

[0089] By integrating and optimizing the spatial clutter intensity distribution map, statistical compliance test report, and key parameters and correction factors of the deduction process, a final electromagnetic clutter deduction result report with a standardized format is generated.

[0090] The core of this step is to integrate key information from the entire simulation process to form a final report with a clear structure and complete information, providing comprehensive simulation results support for subsequent applications (such as radar deployment and electromagnetic environment assessment). The specific implementation method is as follows: The core content of the integration includes four parts: First, an optimized spatial clutter intensity distribution map, which needs to provide a color rendering (including legend, geographic coordinates, and scale), a grayscale rendering (for easy printing), and a core description of the distribution map (such as clutter intensity level classification and distribution patterns of high clutter areas); Second, a statistical conformity test report, which fully includes the test methods, prior data information, test results of individual indicators and overall distribution, anomaly analysis, and test conclusions; Third, key parameters of the extrapolation process, including basic information of the target area (geographical location, area, terrain type), data acquisition parameters (meteorological data acquisition time, frequency), radiation source characteristic parameters (frequency band, polarization mode, transmission power, etc.), computational grid parameters (resolution, dimension, range), and hybrid computational engine parameters (parabolic equation solution step size, empirical scattering model type); Fourth, a set of correction factors, including the values ​​of the multipath propagation attenuation factor α, the scattering cross section adjustment coefficient β, and the environmental adaptation correction coefficient γ, and the calculation basis (nearest neighbor reference area number, weight).

[0091] The final electromagnetic clutter simulation report must adhere to engineering report standards, with a structure comprising six parts: cover, table of contents, abstract, main body, conclusion, and appendix. The cover includes the report title ("Electromagnetic Clutter Simulation Report for Non-Measured Areas"), target area name, simulation time, and compiling unit. The abstract briefly summarizes the simulation's purpose, methods, core results, and verification conclusions. The main body is organized in the following order: "Target Area Overview - Simulation Method Overview - Simulation Parameter Description - Preliminary Simulation Results - Statistical Compliance Verification - Spatial Optimization Results - Correction Factor Explanation." The conclusion clearly states the reliability level of the simulation results, the core clutter distribution patterns, and application recommendations (e.g., radar deployment should avoid high-clutter areas). The appendix contains complete verification data, parameter calculation processes, and excerpts of the original data matrix.

[0092] The core content of the example report includes: The target area overview describes the geographical location as 118°20′–118°30′ E, 30°15′–30°25′ N, covering an area of ​​approximately 306 square kilometers, with terrain mainly consisting of woodland and grassland; the methodology overview briefly describes the entire process of "similar terrain matching – terrain – electromagnetic correlation modeling – correction factor generation – hybrid engine extrapolation – statistical optimization"; the correction factor explanation specifies α=1.063 (weighted calculation based on 5 nearest neighbor reference areas), β=1.105, and γ=1.0006, ensuring the corrected clutter intensity better reflects the actual environment. All data in the report must be labeled with units, and charts must be numbered and explained to ensure readability and traceability. The final generated report is a formatted text file, supporting printing and electronic viewing, and can be directly used in engineering applications.

[0093] Another embodiment of the present invention provides a non-measured region electromagnetic clutter extrapolation system based on similar terrain areas, see [link to relevant documentation]. Figure 3The system may include: The acquisition module 301 is used to acquire geographic information data of the target non-measured area and match one or more reference areas with similar terrain features from the known electromagnetic clutter measurement database; The extraction module 302 is used to extract the measured electromagnetic clutter data and corresponding environmental parameters of each reference area, and to construct a terrain-electromagnetic correlation feature mapping model using a metric-based deep feature learning network. The generation module 303 is used to dynamically generate a correction factor for the multipath propagation and scattering coefficient of the target non-measured area based on the terrain-electromagnetic correlation feature mapping model and in combination with the real-time meteorological parameters and radiation source characteristics of the target non-measured area. The deduction module 304 is used to deduce the spatial clutter intensity distribution of the target non-measured area under a specified frequency band and polarization mode based on the correction factor and through a hybrid calculation engine that couples the parabolic equation and the empirical scattering model. The output module 305 is used to perform statistical characteristic conformity verification and spatial continuity optimization on the spatial clutter intensity distribution, and finally output the electromagnetic clutter inference results for the non-measured area.

[0094] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0095] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0096] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for extrapolating electromagnetic clutter in unmeasured areas based on similar terrain regions, characterized in that, The method includes: Acquire geographic information data of the target non-measured area and match one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database; Measured electromagnetic clutter data and their corresponding environmental parameters for each reference area were extracted, and a terrain-electromagnetic correlation feature mapping model was constructed using a metric-based deep feature learning network. Based on the topographic-electromagnetic correlation feature mapping model, and combined with the real-time meteorological parameters and radiation source characteristics of the target non-measured area, a correction factor for the multipath propagation and scattering coefficient applicable to the area is dynamically generated. Based on the aforementioned correction factor, the spatial clutter intensity distribution of the target's non-measured region under a specified frequency band and polarization mode is deduced through a hybrid computation engine that couples parabolic equations and empirical scattering models. The statistical characteristic conformity test and spatial continuity optimization are performed on the spatial clutter intensity distribution, and the electromagnetic clutter inference results for the non-measured area are finally output.

2. The method according to claim 1, characterized in that, The process of acquiring geographic information data of the target non-measured area and matching one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database includes: Digital elevation model data and land cover type data of the target non-measured area are obtained through a geographic information system to generate a terrain feature dataset of the target area. Multi-scale feature extraction is performed on the terrain feature dataset of the target area, and terrain quantification indicators, including at least terrain roughness, mean slope and standard deviation of elevation, are calculated to generate terrain feature vectors. The similarity between the terrain feature vector and the terrain features of each region in the known electromagnetic clutter measurement database is calculated, and a preliminary list of similar regions is generated by using a metric method based on Euclidean distance. The preliminary list of similar regions is filtered based on a preset terrain similarity threshold, and the final set of matching reference regions is generated by comprehensively considering the area of ​​the region and the completeness of the measured data.

3. The method according to claim 2, characterized in that, The process involves extracting measured electromagnetic clutter data and corresponding environmental parameters for each reference region, and constructing a terrain-electromagnetic correlation feature mapping model using a metric-based deep feature learning network, including: The measured clutter intensity data, meteorological parameters at the corresponding acquisition time, and radar system parameters for each region in the reference region set are extracted from the known electromagnetic clutter measurement database to generate a multimodal dataset of the reference region. The reference region multimodal dataset is normalized and aligned to eliminate differences in the units and collection conditions of data from different regions, thereby generating a standardized reference dataset. A metric-based deep feature learning network is constructed. The terrain feature vector and electromagnetic clutter feature vector from the standardized reference dataset are respectively input into the two branches of the Siamese network to learn the mapping relationship between the two in the shared latent space and generate an initial feature mapping model. The initial feature mapping model is trained using a triplet loss function. The network parameters are optimized to make the electromagnetic features corresponding to similar terrains closer in the latent space, and finally a fully trained terrain-electromagnetic correlation feature mapping model is generated.

4. The method according to claim 3, characterized in that, The step of dynamically generating correction factors for multipath propagation and scattering coefficients applicable to the region based on the topographic-electromagnetic correlation feature mapping model, combined with real-time meteorological parameters and radiation source characteristics of the target non-measured area, includes: Collect real-time meteorological data of the non-measured area of ​​the target, including temperature, humidity and atmospheric pressure, and obtain the operating frequency band, polarization mode and transmission power parameters of the preset radiation source to generate a real-time parameter set of the target area; By fusing the real-time parameter set of the target area with the terrain feature vector, a comprehensive feature description vector of the target area is constructed, and a comprehensive feature vector of the target area is generated. The comprehensive feature vector of the target region is input into the well-trained terrain-electromagnetic correlation feature mapping model. Its representation in the latent space is calculated through forward propagation, and the electromagnetic features of the k nearest reference regions in the latent space are retrieved to generate a set of nearest neighbor electromagnetic features. Based on the nearest electromagnetic feature set, the multipath propagation attenuation factor and scattering cross section adjustment coefficient of the target region relative to the standard theoretical model are calculated by a weighted average algorithm, and finally a set of correction factors applicable to the region is dynamically generated.

5. The method according to claim 4, characterized in that, Based on the correction factor, the spatial clutter intensity distribution of the target's non-measured region under specified frequency bands and polarization modes is deduced using a hybrid computational engine that couples parabolic equations and empirical scattering models, including: The hybrid computing engine is initialized, the standard parabolic equation solver and empirical scattering model library are loaded, and the computing grid is set according to the digital elevation model data of the target area to obtain the initialized computing environment; Input the set of correction factors, the real-time parameter set of the target region, and the characteristic parameters of the radiation source into the hybrid computing engine, configure the boundary conditions of the parabolic equation and the input parameters of the empirical scattering model, and obtain the engine input configuration. The hybrid computing engine is executed first. The parabolic equation module is run to calculate the basic propagation field of electromagnetic waves under complex terrain. Then, the empirical scattering model is called and the correction factor is substituted to calculate the clutter scattering intensity of each grid point, generating a preliminary spatial clutter intensity distribution matrix. The preliminary spatial clutter intensity distribution matrix is ​​post-processed, including unit transformation and logarithmic scaling transformation, to generate a preliminary spatial clutter intensity distribution map of the target non-measured area under a specified frequency band and polarization mode.

6. The method according to claim 5, characterized in that, The statistical characteristic conformity test and spatial continuity optimization of the spatial clutter intensity distribution are performed, and the electromagnetic clutter inference results for the non-measured area are finally output, including: Statistical features are extracted from the preliminary spatial clutter intensity distribution map, and statistics including at least mean, variance, skewness and kurtosis are calculated to generate a distribution statistical feature vector. The statistical feature vector of the distribution is compared with the prior distribution of the statistical features of large sample clutter data in the known measured database. The hypothesis testing method is used to judge the statistical rationality of the inference results and generate a statistical conformity test report. According to the statistical conformity test report, if there are local statistical anomalies, the anisotropic diffusion filtering algorithm is used to smooth the preliminary spatial clutter intensity distribution map, optimize its spatial continuity and suppress unreasonable abrupt changes, and generate an optimized spatial clutter intensity distribution map. By integrating and optimizing the spatial clutter intensity distribution map, statistical compliance test report, and key parameters and correction factors of the deduction process, a final electromagnetic clutter deduction result report with standardized format is generated.

7. A system for extrapolating electromagnetic clutter in non-measured areas based on similar terrain regions, characterized in that, The system includes: The acquisition module is used to acquire geographic information data of the target non-measured area and match one or more reference areas with similar terrain features from a known electromagnetic clutter measurement database; The extraction module is used to extract measured electromagnetic clutter data and their corresponding environmental parameters for each reference area, and to construct a terrain-electromagnetic correlation feature mapping model using a metric-based deep feature learning network. The generation module is used to dynamically generate correction factors for the multipath propagation and scattering coefficients of the target non-measured area based on the terrain-electromagnetic correlation feature mapping model and in combination with the real-time meteorological parameters and radiation source characteristics of the target non-measured area. The deduction module is used to deduce the spatial clutter intensity distribution of the target non-measured area under a specified frequency band and polarization mode based on the correction factor and through a hybrid calculation engine that couples the parabolic equation and the empirical scattering model. The output module is used to perform statistical characteristic conformity verification and spatial continuity optimization on the spatial clutter intensity distribution, and finally output the electromagnetic clutter inference results for non-measured areas.

8. The system according to claim 7, characterized in that, The acquisition module is specifically used for: Digital elevation model data and land cover type data of the target non-measured area are obtained through a geographic information system to generate a terrain feature dataset of the target area. Multi-scale feature extraction is performed on the terrain feature dataset of the target area, and terrain quantification indicators, including at least terrain roughness, mean slope and standard deviation of elevation, are calculated to generate terrain feature vectors. The similarity between the terrain feature vector and the terrain features of each region in the known electromagnetic clutter measurement database is calculated, and a preliminary list of similar regions is generated by using a metric method based on Euclidean distance. The preliminary list of similar regions is filtered based on a preset terrain similarity threshold, and the final set of matching reference regions is generated by comprehensively considering the area of ​​the region and the completeness of the measured data.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images