Multi-source space gamma spectrum fusion and super-resolution reconstruction method and system

Abstract

The application is suitable for the technical field of multi-source data processing, and provides a multi-source space gamma spectrum fusion and super-resolution reconstruction method and system. The method comprises the following steps: target element characteristic peak positioning and peak surface extraction are performed on the preprocessed multi-source spectrum data, and normalization processing is performed based on correlation to obtain a scale-consistent element characteristic peak count value dataset; a point cloud training set with a spatial resolution gradient is constructed based on the element characteristic peak count value dataset; based on a point cloud deep learning network, the point cloud training set is progressively trained in order of low to high spatial resolution to obtain a pre-trained super-resolution reconstruction model; a super-resolution point cloud prediction set is constructed, input into the pre-trained super-resolution reconstruction model for prediction, and the prediction result is evaluated. Through the method, the spectrum data of different sources can be fused and super-resolution reconstructed to obtain an element peak count distribution result higher than the original resolution of any single load.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-source data processing, and particularly relates to a multi-source space gamma spectrum fusion and super-resolution reconstruction method and system. BACKGROUND

[0002] Due to long deep space exploration period, limited number of tasks, and strict load carrying conditions, the acquisition approach of space gamma spectrum itself is extremely limited, especially for celestial bodies outside the Earth. The available data mainly comes from limited orbital exploration tasks. Different exploration tasks have significant differences in satellite flight altitude, orbit design, effective observation time window, and gamma spectrometer detection crystal type, energy resolution capability, etc. The above differences result in inconsistencies in key dimensions such as average sampling number, energy spectrum shape, spatial resolution, and energy resolution of the gamma spectrum obtained by different exploration tasks. It is difficult to directly align and fuse different source gamma spectrums. Therefore, current multi-source space gamma spectrums are difficult to achieve unified element information extraction and joint inversion, so that the energy spectrum data of a single task can only generate an element distribution result matched with its own spatial resolution, which not only fails to fully utilize the limited multi-source spectrum resources, but also limits the spatial accuracy and analysis effect of the final element abundance inversion.

[0003] For the fusion and resolution improvement of gamma spectrum data, existing technologies mainly focus on aerial gamma spectrum and ground exploration scenarios. Related technologies can be broadly divided into two categories: energy spectrum fusion methods and image fusion methods. The research on cross-load and cross-platform fusion of space gamma spectrum is still relatively limited, and there is currently no mature solution that can simultaneously meet the physical consistency and spatial resolution improvement requirements of multi-platform spectrums.

[0004] Energy spectrum fusion methods are usually based on laboratory conditions or fixed measurement environments. For the same target material or the same sample, gamma spectrums are simultaneously acquired by different detectors, and then multiple energy spectrums are fused into a comprehensive energy spectrum output through energy spectrum calibration, energy spectrum broadening correction, deconvolution reconstruction, etc., to improve energy resolution and detection efficiency. However, such methods have the following limitations: first, only single-point or fixed-position energy spectrums are used to construct fusion methods, lacking spatial variables and geographical attributes, and unable to process spatially distributed data for orbital gamma detection; second, laboratory or fixed measurement environments cannot reflect the spatial resolution differences between multi-source space gamma spectrometers caused by factors such as orbital altitude and field of view; third, such methods aim to improve the energy resolution of detectors rather than achieve super-resolution reconstruction of element spatial distribution; fourth, they cannot simultaneously unify the data scale across satellites, orbits, and detectors. Therefore, such fusion technology cannot be used for data fusion and spatial resolution improvement of multi-source space gamma spectrums.

[0005] Image fusion methods typically treat gamma-ray spectrum count maps or element abundance distributions obtained from gamma-ray spectrum inversion as two-dimensional raster data. Algorithms such as multi-source image fusion and super-resolution reconstruction are used to improve image sharpness, signal-to-noise ratio, or detail representation. Common techniques include multi-scale transformation, interpolation enhancement, and deep learning-based image super-resolution. However, these methods have the following drawbacks: First, the fusion object is the mean of all data within the inverted raster; that is, each pixel is only the mean of all gamma-ray spectra within a certain spatial grid, rather than the original gamma-ray spectrum sampling point, ignoring the data characteristics of individual spectra. Second, they cannot utilize the information of individual spectra, failing to leverage the advantages of high sampling volume or high energy resolution data. Third, the fusion process focuses on image texture and spatial structure features, neglecting the physical mechanism of gamma-ray spectrum detection. Fourth, the fusion result is easily dominated by high spatial resolution input, failing to guarantee the contribution weight of high energy resolution or low-error data. In summary, although image fusion methods can improve resolution at the visual level, they cannot fully leverage the advantages of various energy spectrum data and cannot guarantee the physical interpretability and counting accuracy of the fused pixels. Summary of the Invention

[0006] The purpose of this invention is to provide a method for multi-source aerospace gamma spectrum fusion and super-resolution reconstruction, aiming to solve the above-mentioned technical problems.

[0007] This invention is implemented as follows: a method for multi-source aerospace gamma spectrum fusion and super-resolution reconstruction, comprising the following steps:

[0008] Multi-source aerospace gamma spectrum data were acquired and preprocessed to obtain preprocessed multi-source spectrum data.

[0009] The target element feature peaks are located and peak surfaces are extracted from the preprocessed multi-source energy spectrum data, and normalization based on correlation is performed to obtain a dataset of element feature peak counts with consistent scale.

[0010] A point cloud training set with spatial resolution gradient is constructed based on the element feature peak count dataset.

[0011] Based on a point cloud deep learning network, the point cloud training set is progressively trained in order of increasing spatial resolution to obtain a pre-trained super-resolution reconstruction model.

[0012] Construct a super-resolution point cloud prediction set, input it into a pre-trained super-resolution reconstruction model for prediction, and evaluate the prediction results.

[0013] Furthermore, the preprocessing method includes one or more of the following: abnormal signal removal, spectral line drift correction, cosmic ray correction, background correction, and noise correction.

[0014] Further, the step of locating the characteristic peak of the target element and extracting the peak area of the pre-processed multi-source spectral data, and performing normalization based on correlation to obtain a scale-consistent element characteristic peak count value dataset, specifically includes:

[0015] The characteristic peak of the target element is located and the peak area is integrated for the pre-processed multi-source spectral data, and the characteristic peak count value of the target element is obtained from the spectral data to realize the unified conversion from the spectral domain to the element count domain;

[0016] The correlation coefficient between the characteristic peak count values corresponding to the multi-source spectral data is calculated, the spectral data with the maximum total correlation coefficient is selected as the reference source, and then the other sources are normalized to the order of magnitude of the reference source by scaling to obtain a scale-consistent element characteristic peak count value dataset.

[0017] Further, based on the element characteristic peak count value dataset, the step of constructing a point cloud training set with a spatial resolution gradient, specifically includes:

[0018] According to the original spatial resolution of the multi-source spectral data, multiple spatial grid levels are divided from low to high, and for each level of spatial grid, the coordinate position and element count of all sampling points falling into the spatial grid are extracted;

[0019] Each spatial grid and its element count are constructed into a point cloud sample; the average count of the highest spatial resolution data of each spatial grid is set as the supervision target of the point cloud, and the average count of the highest energy resolution data is set as the physical consistency constraint of the point cloud, and the remaining sampling points and their corresponding element counts are used as training feature inputs.

[0020] Further, based on the point cloud deep learning network, the point cloud training set is progressively trained in order of spatial resolution from low to high to obtain a pre-trained super-resolution reconstruction model, specifically including:

[0021] Using a step-by-step migration training strategy from low to high spatial resolution, first train the point cloud sample with low spatial resolution to fit the global trend, then use the trained model as the initial weight, and gradually train and adjust the high spatial resolution point cloud sample data, so that the model gradually learns small-scale spatial features based on maintaining global features, thereby realizing the progressive reconstruction of high-resolution radioactive element distribution features.

[0022] Further, the point cloud deep learning network is a PointNet++ network.

[0023] Further, the step of constructing a super-resolution point cloud prediction set, inputting the pre-trained super-resolution reconstruction model for prediction and result evaluation, and evaluating the prediction result, specifically includes:

[0024] Starting from the spatial grid of the highest original spatial resolution, the spatial grid size is gradually reduced in the direction of higher resolution, and the normalized element characteristic peak count value data set is re-divided according to the spatial grid size to construct a super-resolution point cloud prediction set corresponding to the resolution, and a super-resolution point cloud prediction set is obtained;

[0025] The gradient distribution super-resolution point cloud prediction set is sequentially input into the pre-trained super-resolution reconstruction model, and a prediction result set corresponding to the resolution is output;

[0026] According to the model reconstruction error, spatial smoothness, physical rationality and consistency with low-resolution real data, the prediction result set is comprehensively evaluated, and the best resolution result is selected as the final fusion result under the condition of stable error and statistical reliability, and the super-resolution fusion reconstruction of the multi-source gamma spectrum is realized.

[0027] Another object of the application is to provide a multi-source space gamma spectrum fusion and super-resolution reconstruction system for realizing the multi-source space gamma spectrum fusion and super-resolution reconstruction method.

[0028] The preprocessing module is used for acquiring and preprocessing multi-source space gamma spectrum data to obtain preprocessed multi-source spectrum data;

[0029] The characteristic peak identification and counting module is used for locating and extracting the target element characteristic peak of the preprocessed multi-source spectrum data;

[0030] The correlation coefficient calculation and normalization module is used for normalization processing based on correlation to obtain a scale-consistent element characteristic peak count value data set;

[0031] The point cloud generation module is used for constructing a point cloud training set with gradient spatial resolution based on the element characteristic peak count value data set;

[0032] The multi-scale training module is used for training the point cloud training set in a progressive manner according to the spatial resolution from low to high based on the point cloud deep learning network to obtain a pre-trained super-resolution reconstruction model;

[0033] The super-resolution prediction module is used for constructing a super-resolution point cloud prediction set and inputting the pre-trained super-resolution reconstruction model for prediction;

[0034] The result evaluation and selection module is used for evaluating the prediction result.

[0035] The multi-source space gamma spectrum fusion and super-resolution reconstruction method provided by the application is suitable for a multi-source space gamma spectrum detection scene, in particular, a load combination with different hardware performances of detectors, flight altitudes and orbital coverage densities. Through the method provided by the application, the energy spectrum data collected by different space gamma spectrometers can be fused and super-resolution reconstructed to obtain an element peak count distribution result higher than the original resolution of any single load. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 A flowchart of the multi-source space gamma spectrum fusion and super-resolution reconstruction method provided by the embodiment of the application is shown.

[0037] Figure 2 A specific flowchart of step one in the multi-source space gamma spectrum fusion and super-resolution reconstruction method provided by the embodiment of the application is shown.

[0038] Figure 3 A multi-source element peak linear calibration diagram is shown.

[0039] Figure 4 A specific flowchart of step two in the multi-source space gamma spectrum fusion and super-resolution reconstruction method provided by the embodiment of the application is shown.

[0040] Figure 5 A point cloud distribution state diagram of the same center in multiple scales is shown.

[0041] Figure 6 A specific flowchart of step three in the multi-source space gamma spectrum fusion and super-resolution reconstruction method provided by the embodiment of the application is shown.

[0042] Figure 7 A multi-scale prediction result evaluation diagram is shown. In the diagram, (a) is the average error, (b) is the system error, (c) is the root mean square error, and (d) is the correlation coefficient.

[0043] Figure 8 A comparison diagram of an original resolution element characteristic peak count value image (a) and a super-resolution image (b) is shown.

[0044] Figure 9 An interaction timing diagram of the multi-source space gamma spectrum fusion and super-resolution reconstruction system provided by the embodiment of the application is shown. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the application clearer and more comprehensible, the application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0046] In view of the problems that multi-source space gamma spectrum data is difficult to fuse due to inconsistent average sampling number, spectrum morphology, spatial resolution and energy resolution, an embodiment of the present application provides a multi-source space gamma spectrum fusion and super-resolution reconstruction method based on point cloud deep learning, which can realize fusion and super-resolution reconstruction of multi-source gamma spectrum element counting.

[0047] Specifically, as shown in the figure, Figure 1 In one embodiment of the present application, a multi-source space gamma spectrum fusion and super-resolution reconstruction method is provided, comprising the following steps:

[0048] Step one: obtaining multi-source space gamma spectrum data and preprocessing:

[0049] Obtaining multi-source space gamma spectrum data and preprocessing to obtain preprocessed multi-source spectrum data;

[0050] The preprocessed multi-source spectrum data is subjected to target element characteristic peak positioning and peak surface extraction, and normalized processing based on correlation to obtain a scale-consistent element characteristic peak count value dataset;

[0051] Step two: constructing a point cloud training set with spatial resolution gradient and performing deep learning training:

[0052] Based on the element characteristic peak count value dataset, a point cloud training set with spatial resolution gradient is constructed;

[0053] Based on the point cloud deep learning network, the point cloud training set is progressively trained in order of low to high spatial resolution to obtain a pre-trained super-resolution reconstruction model;

[0054] Step three: constructing a super-resolution point cloud prediction set, inputting the pre-trained super-resolution reconstruction model for prediction, and evaluating the prediction result.

[0055] The embodiment of the present application starts from the original gamma spectrum sampling data, takes multi-source space gamma spectrum detection data as the fusion object, and includes but is not limited to multiple sets of gamma spectrum observation data with differences in spatial resolution, energy resolution, spectrum morphology and sampling density. The purpose of step one of preprocessing multi-source space gamma spectrum data is to correct the physical consistency and extract the characteristics of the original multi-source space gamma spectrum, convert the original spectrum data obtained by different loads into element peak counts under a unified measurement scale, and provide consistent input basis for subsequent point cloud construction and multi-source data fusion.

[0056] As shown in the figure, Figure 2 In a preferred embodiment of the present application, step one specifically comprises the following steps:

[0057] S11, obtaining two or more sets of different source space gamma spectrum data:

[0058] The method provided by the embodiment of the application aims to realize fusion and super-resolution reconstruction among multi-source space gamma spectrums, and therefore needs to introduce gamma spectrum data sets that are different in at least one of source, spatial resolution, energy resolution, spectrum form or sampling density. The more the data sources are, the stronger the model generalization ability is, and the more adaptable the model is to unknown loads and future exploration tasks.

[0059] S12, respectively pre-process each source original gamma spectrum data:

[0060] The original spectrum usually has spectrum line drift, cosmic ray interference, background noise interference and background interference, etc. due to the influence of environmental radiation, extreme climate change, detector component fatigue effect, aging phenomenon and statistical fluctuation, etc. To avoid the propagation of the above errors to the fusion model, the original multi-source space gamma spectrum data collected by different types of instruments need to be pre-processed respectively; the pre-processing specifically includes steps of abnormal signal rejection, spectrum line drift correction, background correction, cosmic ray correction and noise correction, etc. to weaken the influence of instrument stability fluctuation, radiation environment and background noise on the spectrum form, provide consistent input features for element count extraction and subsequent deep learning training, and the specific pre-processing method can be adjusted according to the load performance and background conditions.

[0061] S13, target element characteristic peak positioning and peak count extraction:

[0062] The gamma rays generated by natural radioactive elements are the detection objects of space gamma spectrometers, and the performance differences such as energy resolution and response function of different instruments will cause differences in peak shape and peak height of the same element characteristic peak, and the differentiation degree of adjacent element characteristic peaks is also different. In order to eliminate the interference of spectrum form difference on data fusion, the characteristic peak positioning and peak area integration of target elements are performed on the pre-processed multi-source spectrum data respectively, the characteristic peak count value of the target element is obtained from the spectrum data, and the unified conversion from the spectrum domain to the element count domain is realized.

[0063] S14, calculate the correlation between the element characteristic peak counts of the multi-source spectrums, and perform normalization processing based on the correlation:

[0064] The hardware performance of a gamma spectrometer and the difference in the detection environment determine that the count amplitude of the characteristic peak of the same element measured by different sources is different, which restricts the comparison and fusion of multi-source data. The element abundance in the same region has spatial continuity, so the characteristic peak count of different sources of energy spectrum has a significant positive correlation in space. Specifically, the correlation coefficient between the characteristic peak count values corresponding to the multi-source energy spectrum data is calculated, the energy spectrum data with the maximum total correlation coefficient is selected as the reference source, and then the other sources are mapped to the order of magnitude of the reference source through proportional normalization, to obtain a data set of element characteristic peak count values with consistent scales, so as to eliminate the problem of inconsistent scales of element count amplitudes of different sources, and then improve the fusibility and spatial statistical consistency of multi-source energy spectrum data. Figure 3 For linear calibration of multi-source element peak count, linear regression fitting is performed on the peak count of the reference source and other sources to obtain normalization calibration coefficients, and the count response consistency across sources is realized.

[0065] In the embodiment of the present application, the characteristic peak of the target element of the preprocessed energy spectrum data is located and the characteristic peak count value (element peak count) is extracted, and the energy spectrum under different detectors and different flight conditions is uniformly converted into element peak count form, so as to eliminate the influence of the difference in the detector response function on the spectral line form, and enable the energy spectrum from different sources to be compared in the same count space. Based on the extracted element peak count, a unified spatial grid based on the grid with the lowest spatial resolution is constructed, a pairwise correlation analysis is performed on the count values of different sources, the data source with the highest overall correlation with other sources is selected as the reference, and the count values of other sources are mapped to the reference scale through proportional normalization, so as to eliminate the difference in the sensitivity and energy response of the detectors of different sources, and realize the consistent calibration of the count amplitude across instruments.

[0066] As shown in Figure 4 In a preferred embodiment of the present application, step two specifically includes the following steps:

[0067] S21, dividing gradient grid according to spatial resolution:

[0068] According to the original spatial resolution of the multi-source energy spectrum data, a plurality of spatial grid levels are divided from low to high, for example, from 150km×150km to 50km×50km (the number of levels is determined according to the difference in instrument resolution). In each level of spatial grid, the coordinate position and element count of all sampling points falling within the spatial grid range are extracted to form an element count data set with hierarchical scale structure.

[0069] S22, constructing point cloud training set sample and constraint:

[0070] The sampling points in each spatial grid obtained in the previous step are constructed into a point cloud sample together with the element count thereof. The average count of the highest spatial resolution data of each grid is set as the supervised target of the point cloud, ensuring that the prediction accuracy during model training has physical meaning; the average count of the highest energy resolution data is set as the physical consistency constraint of the point cloud, ensuring that the spectral peak information and element distinguishing energy spectrum are not lost, and the advantage that high energy resolution data can accurately separate element characteristic peaks is exerted; and the remaining sampling points and the corresponding element count thereof are taken as the training feature input, so that the point cloud training set simultaneously maintains spatial hierarchical features and energy spectrum reliability.

[0071] In the embodiment of the present application, a point cloud training set with increasing resolution is constructed using the normalized full element characteristic peak count value data set. Specifically, starting from the lowest resolution spatial grid scale, according to the spatial coverage capability and resolution level of various observation data, several (such as 5-10 levels) spatial scales are divided, and all observation point coordinates in each spatial grid are encapsulated into a point cloud sample together with the corresponding element count. For the point cloud of each scale, the average count of the detector with the highest spatial resolution in each grid is taken as the supervised target, and the average count of the detector with the highest energy resolution is taken as the physical consistency constraint value, which is introduced into the model loss function, so as to take into account both high spatial resolution information and high energy spectrum accuracy information. In actual application, the point cloud distribution state of the same center at multiple scales is as shown in Figure 5 .

[0072] S23, sequentially input the point cloud training set in order from low to high resolution into the point cloud deep learning network for training:

[0073] The sampling principle of the space gamma spectrometer determines that the sampling points are discretely distributed, and the sampling point density and total data volume of different sources are also different, and the spatial features conform to the feature expression of the point cloud. The point cloud deep learning network can directly learn based on the discrete sampling points, without regular gridding or interpolation, and can extract the spatial correlation features, geometric distribution structure and statistical mode of the sampling points in the region in the original spatial coordinate system, so as to adapt to the sampling characteristics of the space gamma detection data and retain the authenticity and integrity of the spatial energy spectrum information. The embodiment of the present application is based on the PointNet++ network, uses a step-by-step migration training strategy from low to high resolution, first trains with a large grid (low spatial resolution point cloud) to fit the global trend, then uses the trained model as the initial weight, and gradually trains and adjusts the small grid (high spatial resolution point cloud) data, so that the model gradually learns the small-scale spatial features on the basis of maintaining the global features, thereby realizing the progressive reconstruction of the high-resolution radioactive element distribution features, and obtaining a pre-trained super-resolution reconstruction model.

[0074] In the embodiment of the present application, the model training adopts a low-to-high resolution incremental strategy, takes the lowest resolution point cloud data as input, uses a point cloud deep learning network (such as PointNet++ network) for training, and after the training converges, takes the model parameters as the initial state of the next resolution level, and gradually inputs higher resolution point cloud data for training. Through the multi-resolution training mechanism of progressive levels, the model learns the multi-scale spatial organization structure while gradually acquiring high-resolution spectral details and physical constraint characteristics, and finally obtains a pre-trained super-resolution reconstruction model that can maintain physical consistency and statistical stability at a high resolution scale.

[0075] As shown in Figure 6 In a preferred embodiment of the present application, step three specifically includes the following steps:

[0076] S31, constructing a super-resolution point cloud prediction set:

[0077] Starting from the highest original spatial resolution spatial grid, gradually reduce the spatial grid size in the direction of higher resolution (for example, 50km, 45km, 40km, … 5km, make sure that the number of sampling points in each spatial grid meets the minimum data requirement of point cloud learning to ensure statistical stability and training effectiveness), redivide the pre-processed normalized element feature peak count value data according to the spatial grid size, construct a super-resolution point cloud prediction set of corresponding resolution (for reference the construction method of the above point cloud training set), and use it to calculate high-resolution element count prediction values step by step, and realize resolution extrapolation.

[0078] S32, input the point cloud prediction set into the pre-trained super-resolution reconstruction model for step-by-step prediction, and evaluate the fusion results of each level:

[0079] The gradient distribution of the super-resolution point cloud prediction set is input into the pre-trained super-resolution reconstruction model in turn, and the prediction result set of the corresponding resolution is output. Then, according to the model reconstruction error, spatial smoothness, physical reasonableness and consistency with low-resolution true data, the prediction result set is comprehensively evaluated, and the best resolution result is selected as the final fusion result under the condition of stable error and statistical credibility, realizing the super-resolution fusion reconstruction of multi-source gamma spectrum.

[0080] In the model inference stage, the normalized unified count data is taken as input, a prediction grid with higher resolution than that in the training stage is further constructed based on the highest original resolution space grid, a data set with gradually increased spatial resolution is formed, and the trained model is input for prediction. The model output is the count value of the super-resolution element corresponding to each space grid, so as to obtain the element distribution reconstruction result of the high-resolution scale. Through the evaluation of the prediction results of different resolutions, the prediction set with the highest spatial resolution under the premise of ensuring the consistency of the count and the reasonable controllable error is selected as the final fusion result, so as to realize the high-precision fusion and super-resolution reconstruction of the multi-source space gamma-ray spectrum data.

[0081] In actual application, the result predicted by the above model is evaluated as shown in Figure 7 The evaluation indexes used are respectively: root mean square error, used for measuring the error amplitude and sensitive to large error; average error, used for reflecting the average error level; system error, used for representing the systematic overestimation or underestimation trend of the model; and correlation coefficient, used for measuring the consistency of the prediction result and the spatial distribution of the original data. As shown in Figure 7 When the super-resolution grid is further refined from 30 km to 20 km, the system error will significantly increase and the spatial correlation will obviously decrease, indicating that excessive refinement will introduce uncertainty, therefore, the grid of 30 km is taken as the best super-resolution output scale in the region.

[0082] In addition, the comparison of the super-resolution element feature peak count value image obtained by fusing the multi-source spectrum data and the original image is as shown in Figure 8 A typical region (the region size is about 2000 km x 2000 km) with significant geochemical gradient characteristics is selected, the original data with a spatial resolution of about 50 km is compared with the super-resolution result of about 30 km obtained by the embodiment of the present application. The super-resolution image is significantly better than the original image in element gradient enhancement, geochemical anomaly extraction and detail expression, proving that the embodiment of the present application can obtain element distribution information with higher spatial resolution while ensuring data reliability.

[0083] As shown in Figure 9 In another embodiment of the present application, a multi-source space gamma-ray spectrum fusion and super-resolution reconstruction system is also provided, which is used to realize the multi-source space gamma-ray spectrum fusion and super-resolution reconstruction method described above, and comprises:

[0084] A preprocessing module is configured to acquire multi-source space gamma-ray spectrum data and perform preprocessing to obtain preprocessed multi-source spectrum data.

[0085] A feature peak identification and counting module is configured to locate the target element feature peak and extract the peak surface of the preprocessed multi-source spectrum data.

[0086] a correlation coefficient calculation and normalization module configured to perform normalization based on the correlation to obtain a scale-consistent element characteristic peak count value dataset;

[0087] a point cloud generation module configured to construct a point cloud training set with a spatial resolution gradient based on the element characteristic peak count value dataset;

[0088] a multi-scale training module configured to perform progressive training on the point cloud training set in an order from low to high spatial resolution based on a point cloud deep learning network to obtain a pre-trained super-resolution reconstruction model;

[0089] a super-resolution prediction module configured to construct a super-resolution point cloud prediction set and input the pre-trained super-resolution reconstruction model for prediction;

[0090] a result evaluation and selection module configured to evaluate the prediction result.

[0091] Figure 9 An interactive timing diagram of the system provided by the embodiment of the present application shows the internal module calling sequence and data flow path of the system, mainly including the interactive process of multi-source data acquisition, preprocessing, peak count extraction, cross-source normalization calibration, point cloud construction, multi-resolution progressive training, super-resolution prediction and result evaluation and selection.

[0092] As shown in Figure 9 , the system first calls original spectrum data from a multi-source space gamma spectrum database and submits the data to a preprocessing module for spectrum line drift correction, background subtraction and noise correction, etc. The preprocessed spectrum data enters a characteristic peak identification and counting module to extract characteristic peak count values corresponding to target nuclides. Then, a correlation coefficient calculation and normalization module performs correlation analysis on peak counts from different sources, determines a reference source and performs cross-source peak count normalization calibration to obtain a scale-consistent count dataset. A point cloud generation module constructs a multi-level point cloud dataset according to different spatial resolution grid sizes and sequentially passes the dataset to a multi-scale training module to perform progressive training from low to high resolution. The optimal model after training and the super-resolution point cloud prediction set are passed to a super-resolution prediction module together, which outputs higher resolution element count prediction layers step by step. Finally, a result evaluation and selection module filters the optimal resolution output based on error consistency, spatial smoothness and physical reasonableness, etc.

[0093] In a preferred embodiment of the present application, the characteristic peak identification and counting module specifically includes:

[0094] a spectrum characteristic peak positioning unit configured to construct a scale relationship between channel address and energy and identify characteristic peak positions corresponding to target radioactive elements;

[0095] The feature peak count value extraction unit is configured to obtain the feature peak count value of the target element by peak fitting and peak area integration, and extract the energy value generated by the corresponding element from the entire spectrum.

[0096] The correlation coefficient calculation and normalization module specifically includes:

[0097] The peak count correlation analysis unit is configured to calculate the element feature peak count correlation of the multi-source energy spectrum in the same region and spatial resolution, and determine the data source with the highest correlation sum as the reference source;

[0098] The cross-source normalization calibration coefficient calculation unit is configured to establish a metrological mapping relationship between the peak counts of other sources and the reference source, and solve the normalization calibration coefficients of each source;

[0099] The peak count normalization and output unit is configured to normalize the peak counts of each source to the reference source scale according to the calibration coefficients, and output the multi-source peak count data in a unified scale, thereby providing input data for point cloud sample construction and multi-scale training.

[0100] The point cloud generation module specifically includes:

[0101] The multi-scale spatial division unit is configured to divide multi-level spatial grids from low to high according to the native resolution of the multi-source data, thereby forming a resolution-gradual hierarchical set (e.g., 150 km to 100 km to 50 km…), and enabling the model to learn the spatial distribution features at different scales step by step;

[0102] The point cloud sample construction and constraint generation unit is configured to divide and construct the normalized peak counts in each spatial grid range into point cloud samples, and take the count mean calculated in the point cloud according to the highest spatial resolution as a supervision label, and take the count mean of the highest energy resolution as a physical consistency constraint label, thereby providing a structured training target for model learning.

[0103] The multi-scale training module specifically includes:

[0104] The multi-scale point cloud training unit is configured to train the model step by step in order from low to high spatial resolution, and input point cloud samples at each scale to extract spatial structure features at the corresponding scale;

[0105] The step-by-step weight inheritance unit is configured to take the model parameters obtained by training at the previous scale as the initial weights for training at the next scale, thereby enabling the model to further learn high-resolution details on the basis of maintaining low-resolution stability, and improving the cross-scale generalization ability;

[0106] The loss optimization and model convergence control unit is configured to supervise the loss function in model training, and realize stable convergence of the model by combining early stopping, dynamic constraint weight adjustment, and other methods;

[0107] An optimal model output unit is configured to save an optimal model obtained by a multi-scale point cloud training process, and provide a deep model support for a subsequent super-resolution point cloud prediction stage.

[0108] The super-resolution prediction module specifically includes:

[0109] A super-resolution multi-scale space grid division unit is configured to start from an original highest spatial resolution and gradually refine the space grid based on resolution requirements;

[0110] A predicted point cloud construction unit is configured to divide and construct the peak count data in each space grid range into a point cloud prediction set;

[0111] A point cloud network inference unit is configured to input each point cloud prediction set into a point cloud model optimized by multi-scale training, and realize a step-by-step super-resolution prediction of element peak count.

[0112] A multi-scale prediction result generation unit is configured to output a prediction result grid data set corresponding to the spatial resolution grid.

[0113] The result evaluation and selection module is specifically configured to comprehensively evaluate the multi-scale prediction results according to error indicators, spatial consistency, element distribution characteristics, etc. Finally, the optimal super-resolution element count output module generates a super-resolution element peak count distribution fused with multi-source observations, and provides support for subsequent mapping, scientific analysis and data products.

[0114] It should be noted that each of the above modules or units can be implemented in the form of a computer program, which can be run on a computer device, and the computer program can be stored in the memory of the computer device to enable the processor to execute each step of the above method.

[0115] It should be understood that although each step in the flowchart of each embodiment of the present application is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise stated herein, the execution of these steps has no strict sequence limitation, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least a part of other steps or sub-steps or stages of other steps.

[0116] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the program can include the processes of the above-mentioned embodiments of each method. Any reference to memory, storage, database or other medium used in each embodiment provided by the present application can include non-volatile and / or volatile memory.

[0117] The above embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but should not be understood as a limitation on the patent scope of the present application. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for multi-source aerospace gamma spectrum fusion and super-resolution reconstruction, characterized in that, Includes the following steps: Multi-source aerospace gamma spectrum data were acquired and preprocessed to obtain preprocessed multi-source spectrum data. The target element feature peaks are located and peak surfaces are extracted from the preprocessed multi-source energy spectrum data, and normalization based on correlation is performed to obtain a dataset of element feature peak counts with consistent scale. A point cloud training set with spatial resolution gradient is constructed based on the element feature peak count dataset. Based on a point cloud deep learning network, the point cloud training set is progressively trained in order of increasing spatial resolution to obtain a pre-trained super-resolution reconstruction model. Construct a super-resolution point cloud prediction set, input it into a pre-trained super-resolution reconstruction model for prediction, and evaluate the prediction results; The steps involved in locating and extracting the characteristic peaks of the target elements from the preprocessed multi-source energy spectrum data, followed by correlation-based normalization to obtain a dataset of element characteristic peak counts with consistent scale, specifically include: The characteristic peaks of the target elements are located and the peak area is integrated for the preprocessed multi-source energy spectrum data. The characteristic peak count values ​​of the target elements are obtained from the energy spectrum data, realizing a unified conversion from the energy spectrum domain to the element counting domain. Calculate the correlation coefficient between the characteristic peak counts corresponding to the multi-source energy spectrum data, select the energy spectrum data with the largest sum of correlation coefficients as the reference source, and then map the other sources to the order of magnitude of the reference source through proportional normalization to obtain a dataset of element characteristic peak counts with consistent scale. The steps for constructing a point cloud training set with spatial resolution gradient based on an element feature peak count dataset include: Based on the original spatial resolution of the multi-source energy spectrum data, multiple spatial grid levels are divided from low to high. For each level of spatial grid, the coordinate positions and element counts of all sampling points falling within that spatial grid are extracted. Each spatial grid is used to construct a point cloud sample by combining the sampling points and their element counts. The average count of the highest spatial resolution data in each spatial grid is set as the supervision target of the point cloud, and the average count of the highest energy resolution data is set as the physical consistency constraint of the point cloud. The remaining sampling points and their corresponding element counts are used as training feature inputs.

2. The multi-source aerospace gamma spectrum fusion and super-resolution reconstruction method according to claim 1, characterized in that, The preprocessing methods include one or more of the following: abnormal signal removal, spectral line drift correction, cosmic ray correction, background correction, and noise correction.

3. The multi-source aerospace gamma spectrum fusion and super-resolution reconstruction method according to claim 1, characterized in that, The steps for obtaining a pre-trained super-resolution reconstruction model based on a point cloud deep learning network, where the point cloud training set is trained progressively according to increasing spatial resolution, include: Using a stepwise transfer training strategy with spatial resolution from low to high, the model is first trained with low spatial resolution point cloud samples to fit the global trend. Then, the trained model is used as the initial weights to train and adjust with high spatial resolution point cloud sample data step by step. This allows the model to gradually learn small-scale spatial features while maintaining global features, thereby achieving a progressive reconstruction of high-resolution radioactive element distribution features.

4. The multi-source aerospace gamma spectrum fusion and super-resolution reconstruction method according to claim 3, characterized in that, The point cloud deep learning network is the PointNet++ network.

5. The multi-source aerospace gamma spectrum fusion and super-resolution reconstruction method according to claim 1, characterized in that, The steps for constructing a super-resolution point cloud prediction set, inputting it into a pre-trained super-resolution reconstruction model for prediction, and evaluating the prediction results include: Starting with the highest original spatial resolution spatial grid, the spatial grid size is gradually reduced towards higher resolutions. The normalized element feature peak count dataset is then re-divided according to the spatial grid size to construct a super-resolution point cloud prediction set corresponding to the resolution. The super-resolution point cloud prediction set with gradient distribution is sequentially input into the pre-trained super-resolution reconstruction model, and the corresponding resolution prediction result set is output. The prediction result set is comprehensively evaluated based on the model reconstruction error, spatial smoothness, physical rationality, and consistency with low-resolution real data. Under the condition of stable error and statistical reliability, the best resolution result is selected as the final fusion result, so as to realize the super-resolution fusion reconstruction of multi-source gamma spectrum.

6. A multi-source aerospace gamma spectrum fusion and super-resolution reconstruction system, used to implement the multi-source aerospace gamma spectrum fusion and super-resolution reconstruction method according to any one of claims 1-5, characterized in that, include: The preprocessing module is used to acquire multi-source aerospace gamma spectrum data and perform preprocessing to obtain preprocessed multi-source spectrum data; The feature peak identification and counting module is used to locate the feature peaks of target elements and extract the peak surfaces of preprocessed multi-source energy spectrum data. The correlation coefficient calculation and normalization module is used to perform normalization based on correlation to obtain a dataset of element feature peak counts with consistent scale. The point cloud generation module is used to construct a point cloud training set with spatial resolution gradient based on the element feature peak count dataset. The multi-scale training module is used to progressively train the point cloud training set according to the spatial resolution from low to high based on the point cloud deep learning network, so as to obtain the pre-trained super-resolution reconstruction model. The super-resolution prediction module is used to construct a super-resolution point cloud prediction set and input it into a pre-trained super-resolution reconstruction model for prediction. The results evaluation and selection module is used to evaluate the prediction results.

Images