Two-stage multi-source data fusion geological mapping method based on U-Net network

Abstract

The present invention discloses a two-stage multi-source data fusion geological mapping method based on U-Net network, comprising: obtaining multi-source geological data of the area to be mapped; pre-processing the multi-source geological data; and using U-Net neural network to fusion the basic geological map of the entire area to be mapped. total The method performs a coarse classification; uses a U-Net neural network to subdivide the first-stage coarse-classified intelligent mapping results; integrates the second-stage results with the first-stage results to obtain an intelligent mapping result for the entire area to be mapped; conducts field verification on the intelligent mapping results for the entire area to obtain a revised intelligent mapping result; and completes a geological interpretation map based on the revised intelligent mapping result. The present invention has the beneficial effects of utilizing a relatively small training area and accelerating the prediction of the area to be mapped in a GPU environment, achieving high resolution and generalization capabilities, and providing guidance for mapping covered areas and restricted work areas.

Description

Technical Field

[0001] The present invention relates to the technical field of geological mapping, and in particular to a two-stage multi-source data fusion geological mapping method based on a U-Net network. Background Art

[0002] Geological mapping is a fundamental method in mineral prospecting and exploration. However, due to factors such as climate, topography, and legal regulations, field geological surveys are often limited and consume significant manpower, material, and financial resources. Furthermore, the collected data is often incomplete, and geological data that cannot be collected manually directly impacts the results of geological mapping. Therefore, applying modern information technology to geological surveys and leveraging deep learning for intelligent geological mapping is crucial.

[0003] The current geological mapping methods include: (1) Surface geological survey: using route geological survey and measured profile measurement as the main means, analyzing previous research data and conducting surveys. (2) Remote sensing geological interpretation: using the topography, water system and geological body's schlieren, color tone, roughness, etc. as interpretation signs to trap geological bodies, summarize the remote sensing image characteristics of different geological units to compile remote sensing interpretation geological maps. (3) Geological mapping based on statistics, including: ① Inversion of aeromagnetic data: based on the magnetic differences between rocks, magnetic units are divided through cluster analysis and other methods to provide a basis for the establishment of geological units for geological mapping. ② Soil geochemical inversion: using the inheritance of chemical composition between soil components and underlying rocks, the composition characteristics of the underlying bedrock are inverted through soil components to provide a chemical basis for geological mapping. (4) Geological mapping methods based on machine learning, such as using machine learning methods such as random forests and support vector machines to classify the geological structure of a given target area.

[0004] Deep learning (DL) is a key branch of machine learning and a new research direction in artificial intelligence. Its motivation is to build and simulate neural networks that analyze and learn like the human brain. It has been successfully applied to image classification, semantic segmentation, object detection, and image generation, and is driving significant progress in artificial intelligence and human-computer interaction. Deep learning networks require no parameter tuning after training and can rapidly reconstruct large amounts of data using GPUs, a feat that is difficult for most existing algorithms. Regarding the application of DL to geological mapping, Chen et al. combined Sentinel-2 and ASTER data to map seven lithologies in the Mount Isa area using a convolutional neural network (CNN), achieving a prediction accuracy of 75%. This method first spatially resampled multi-sensor remote sensing data and geophysical data from different bands and fused them as model input. The network was then trained and each pixel in the multi-band image was classified as a rock type based on a defined probability threshold. However, this method suffers from a limited number of predicted categories, a significantly larger training area than the target area, and low resolution, resulting in inaccurate predictions in many areas. Fu Si et al. proposed a method for geological body identification that uses multimodal feature fusion. Combining geophysical, geochemical, and remote sensing imagery, they used deep neural networks (DNNs) and CNNs to extract features from each modality. These features were then combined to create multimodal features. Finally, a fully connected neural network was used to fuse these features for geological body classification. However, research in shallow overburden areas revealed that the classification of some geological bodies in the predicted image conflicted with the spatial location of the measured image.

[0005] These research results demonstrate the enormous potential of DL in geological mapping, but they also expose a significant challenge: the large training sample requirements and the difficulty of generalizing and predicting the network. Effective network testing requires a large amount of labeled training data. Without sufficient training data, neural networks generally exhibit poor generalization performance. Summary of the Invention

[0006] In order to address the above technical problems, the present invention provides a two-stage multi-source data fusion geological mapping method based on the U-Net network. This method fully exploits the relationship between gamma spectroscopy data and other geophysical data and different geological types, takes into account the characteristics of overall and partial mapping, and significantly reduces the cost of field geological exploration. By extracting features through the network, classification prediction of geological bodies can be completed in the presence of vegetation coverage. The algorithm proposed in the present invention has high resolution and generalization ability on the 1:50,000 Liyutang map in Hunan Province. The method includes the following steps:

[0007] S1. Obtain multi-source geological data of the area to be mapped, and record the basic geological map of the area to be mapped as Ω total ;

[0008] S2. preprocessing the multi-source geological data to obtain preprocessed multi-source geological data;

[0009] S3, using U-Net neural network to map the basic geological map of the entire area to be mapped total Perform coarse classification to obtain the first stage of coarse classification intelligent mapping results;

[0010] S4, using the U-Net neural network to subdivide the first stage coarse intelligent mapping results to obtain the second stage subdivided intelligent mapping results;

[0011] S5. Integrate the second-stage subdivided intelligent mapping results into the first-stage coarse-divided intelligent mapping results to obtain the basic geological map Ω of the entire area to be mapped. total Intelligent mapping results;

[0012] S6. Basic geological map of the entire area to be mapped Ω total The intelligent mapping results are verified in the field to obtain the revised intelligent mapping results;

[0013] S7. Complete the geological interpretation map based on the corrected intelligent mapping results.

[0014] Furthermore, step S3 is specifically as follows:

[0015] S31: Get training set S input_stage1 , and the training set S input_stage1 The corresponding predicted label is recorded as S label_stage1 ;

[0016] S32: Constructing a U-Net neural network for coarse geological type prediction;

[0017] S33: Using training set S input_stage1 and the corresponding predicted label S label_stage1 , train the U-Net neural network for coarse geological type prediction, obtain the prediction loss, minimize the loss and obtain the optimized network parameters

[0018] S34: Based on the optimized network parameters Obtain the trained U-Net neural network for coarse geological type prediction;

[0019] S35: Input the pre-processed multi-source geological data of the entire area to be mapped into the trained coarse geological type prediction U-Net neural network to obtain the first stage coarse classification intelligent mapping result, so that the basic geological map of the entire area to be mapped Ω total Divided into n types of coarse-grained regions Ω k (k=1,2,...,n).

[0020] Furthermore, step S4 is specifically as follows:

[0021] S41: Get the training set and classify the regions into n types according to Ω k (k=1,2,...,n) Randomly cut the preprocessed multi-source geological data corresponding to the training set into multiple N×N sub-blocks to obtain n sub-training sets The corresponding predicted label is recorded as

[0022] S43: Constructing n sub-segment geological type prediction U-Net neural network;

[0023] S44: Using n sub-training sets, train n sub-segmented geological type prediction U-Net neural networks one by one, obtain prediction loss, and minimize the loss to obtain optimized network parameters

[0024] S45: Based on the optimized network parameters Obtain the trained n sub-segment geological type prediction U-Net neural network;

[0025] S46: n types of roughly divided regions Ω k The multi-source data corresponding to (k=1,2,...,n) are respectively input into the corresponding trained n sub-segment geological type prediction U-Net neural network to obtain the intelligent mapping results of the second stage of segmentation

[0026] The optimized network parameters involved in step S33 and step S44 Its general expression formula is as follows:

[0027]

[0028] Among them, F is the coarse geological type prediction U-Net neural network or one of the n sub-segmented geological type prediction U-Net neural networks, S label For S label_stage1 or x∈Φ, is the data point in the training set; the mapping l:Φ→{1,2,...,M} represents the true category corresponding to each data point.

[0029] The beneficial effects provided by the present invention are:

[0030] The U-Net network-based two-stage multi-source data fusion geological mapping method of the present invention utilizes the collected multi-source data and uses a deep learning method to automatically extract the characteristics of various data, fully explore the relationship between the data and different geological types, and adopts a two-stage strategy to complete intelligent geological mapping, assisting geological experts in carrying out geological surveys.

[0031] The algorithm and implementation process proposed in this invention can utilize a smaller training area and accelerate the prediction of the area to be mapped in a GPU environment. Experimental results on the 1:50,000 Liyutang map of Hunan Province show that the algorithm has high resolution and generalization ability, and has certain guiding significance for the mapping work of covered areas and restricted working areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic flow chart of the method of the present invention.

[0033] Figure 2 It is a schematic diagram of the algorithm framework of the geological mapping method of the present invention based on the two-stage multi-source data fusion of the U-Net network.

[0034] Figure 3 This is a 1:50,000 basic geological map of Liyutang, Hunan Province.

[0035] Figure 4 These are the airborne gamma-ray spectrum data of the Liyutang area, where (a) is the total amount of radioactive gamma-ray spectrum data, denoted as γ; (b) is the magnetotelluric data; and (c) is the gamma-ray spectrum and magnetotelluric interpretation lines.

[0036] Figure 5 The processed and transformed aeromagnetic anomaly data include (a) the ΔT-transformed post-pole magnetic anomaly, denoted as MA; (b) the ΔT-transformed magnetic anomaly 2 km above the pole, denoted as RF; and (c) the third-order details in the wavelet multi-scale analysis of aeromagnetic data, denoted as third-order details.

[0037] Figure 6 This is the first stage of rough classification, where (a) is the 5-category rough classification label of the carp pond; (b) is the 5-category rough classification result.

[0038] Figure 7 This is the second stage of segmentation, where (a) is the 26-category label of the carp pond; (b) is the 26-category segmentation result.

[0039] Figure 8 These are the results of field occurrence point verification, where (a) is the verification of coarse occurrence points, and (b) is the verification of fine occurrence points.

[0040] Figure 9It is an intelligent geological interpretation map drawn by experts based on the intelligent mapping results. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0042] Please refer to Figure 1 , Figure 1 The figure is a flow chart of the method of the present invention; the two-stage multi-source data fusion geological mapping method based on the U-Net network includes the following:

[0043] S1. Obtain multi-source geological data of the area to be mapped, and record the basic geological map of the area to be mapped as Ω total ;

[0044] It should be noted that the process of preparing multi-source geological data is as follows:

[0045] Collect existing geological basic data, geophysical data, geochemical data, remote sensing data, as well as publicly published papers and test analysis data in the area to be mapped.

[0046] S2. preprocessing the multi-source geological data to obtain preprocessed multi-source geological data;

[0047] It should be noted that the data preprocessing process is as follows:

[0048] The collected multi-source geological data are processed to unify data format and size, standardize attribute meaning, declassify confidential data, and perform transformations such as rotation, analytical extension, magnetic pole transformation, and wavelet multi-scale analysis. Experts interpret and label the data.

[0049] S3, using U-Net neural network to map the basic geological map of the entire area to be mapped total Perform coarse classification to obtain the first stage of coarse classification intelligent mapping results;

[0050] It should be noted that step S3 is specifically as follows:

[0051] S31: Get training set S input_stage1 , and the training set S input_stage1 The corresponding predicted label is recorded as S label_stage1 ;

[0052] It should be noted that in the actual processing process, firstly, Ω totoal After human intervention, n types are roughly divided, denoted as Ω stage1 However, the types here are divided by human intervention rather than predicted by the model.

[0053] Here the coarse-divided area training set S input_stage1, which also corresponds to the roughly divided n types, and is randomly cropped into n N×N sub-blocks.

[0054] S32: Constructing a U-Net neural network for coarse geological type prediction;

[0055] S33: Using training set S input_stage1 and the corresponding predicted label S label_stage1 , train the U-Net neural network for coarse geological type prediction, obtain the prediction loss, minimize the loss and obtain the optimized network parameters

[0056] S34: Based on the optimized network parameters Obtain the trained U-Net neural network for coarse geological type prediction;

[0057] S35: Input the pre-processed multi-source geological data of the entire area to be mapped into the trained coarse geological type prediction U-Net neural network to obtain the first stage coarse classification intelligent mapping result, so that the basic geological map of the entire area to be mapped Ω total Divided into n types of coarse-grained regions Ω k (k=1,2,...,n).

[0058] S4, using the U-Net neural network to subdivide the first stage coarse intelligent mapping results to obtain the second stage subdivided intelligent mapping results;

[0059] It should be noted that step S4 is specifically as follows:

[0060] S41: Obtain training data and classify the regions into n types according to Ω k (k=1,2,...,n) Randomly cut the preprocessed multi-source geological data corresponding to the training data into multiple N×N sub-blocks to obtain n sub-training sets The corresponding predicted label is recorded as

[0061] S43: Constructing n sub-segment geological type prediction U-Net neural network;

[0062] S44: Using n sub-training sets, train n sub-segmented geological type prediction U-Net neural networks one by one, obtain prediction loss, and minimize the loss to obtain optimized network parameters

[0063] S45: Based on the optimized network parameters Obtain the trained n sub-segment geological type prediction U-Net neural network;

[0064] S46: n types of roughly divided regions Ωk The multi-source data corresponding to (k=1,2,...,n) are respectively input into the corresponding trained n sub-segment geological type prediction U-Net neural network to obtain the intelligent mapping results of the second stage of segmentation

[0065] It should be noted that the optimized network parameters involved in step S33 and step S44 are Its general expression formula is as follows:

[0066]

[0067] Among them, F is the coarse geological type prediction U-Net neural network or one of the n sub-segmented geological type prediction U-Net neural networks, S label For S label_stage1 or x∈Φ, is the data point in the training set; the mapping l:Φ→{1,2,...,M} represents the true category corresponding to each data point. In the first stage of coarse classification, Ω total Contains n categories, that is, M = n; in the second stage of subdivision, Ω k Contains k geological type, that is, M = t k . is the output of the network, which indicates the probability that the data point x belongs to the Mth geological type. The optimized network parameter θ F Make the probability that the geological type of the network-predicted data point x is the true category l(x) close to 1.

[0068] S5. Integrate the second-stage subdivided intelligent mapping results into the first-stage coarse-divided intelligent mapping results to obtain the basic geological map Ω of the entire area to be mapped. total Intelligent mapping results;

[0069] Ω total When performing coarse segmentation, the prediction result p of each channel of the U-Net network is (m) See formula:

[0070]

[0071] where Φ m (m=1,2,...,n), indicating different geological types.

[0072] Ω k , the prediction result p of each channel of the U-Net network during the training process (m) See formula:

[0073]

[0074] Among them, Φ km (k=1,2,..,n;m=1,2,...,t k ) represent different geological types in different areas.

[0075] It should be noted that the optimized network parameters θ F Finally, the network performance is tested on the area to be mapped, and the overall accuracy (ACC) is used to evaluate the network performance. The formula is as follows:

[0076]

[0077] Among them, Num correct The number of correct pixels predicted by the network, Num all For S label The total number of pixels in the network is calculated, and the model with the highest accuracy is used as the final prediction network.

[0078] S6. Basic geological map of the entire area to be mapped Ω total The intelligent mapping results are verified in the field to obtain the revised intelligent mapping results;

[0079] It should be noted that step S6 specifically involves: total Based on the intelligent mapping results, 3 to 5 verification routes that cross the major boundary lines of the geological body are designed to control the boundaries of the regional geological body, optimize the accuracy of boundary recognition, and obtain the corrected intelligent mapping results.

[0080] S7. Complete the geological interpretation map based on the corrected intelligent mapping results.

[0081] It should be noted that step S7 is specifically to assist in outlining the geological boundaries based on the corrected intelligent mapping results, on the basis of comprehensive various geological data, according to the stratigraphic occurrence and topography, and according to the V-shaped rule, to obtain a geological interpretation map that has the identification content of geological unit boundaries (tectonic lines, classification unit boundaries, deformation patterns, etc.) and special geological bodies (strata, rock bodies, etc.) as a whole.

[0082] The entire U-Net network structure of this application is further explained below.

[0083] Please refer to Figure 2 ,like Figure 2As shown in the figure, the two-stage strategy of the present invention is composed of multiple U-Net networks in parallel, where the network for the coarse division stage is network 1, and the networks for the fine division stage are networks 2, 3, ..., n+1. Intelligent geological mapping experiments are carried out separately using multi-source data that can reflect the geological type characteristics of each region, and the loss function is the cross entropy loss between the network output and the label.

[0084] The present invention conducts a two-stage experiment based on the U-Net network and fuses multi-source data in the input layer to form a multi-channel input.

[0085] In the first coarse classification phase, similar geological types are first classified based on basic geological data, resulting in n coarse geological types. Network 1 is then trained to perform coarse classification predictions for the area to be mapped, resulting in intelligent geological mapping results for the coarse phase. In the second subdivision phase, networks 2, 3, ..., n+1 are trained on each of the n coarse regions, performing subdivision predictions for the area to be mapped, resulting in intelligent geological mapping results for the subdivision phase. Finally, the subdivision mapping results are integrated with the coarse mapping results to obtain an intelligent geological mapping result for the entire area to be mapped. Based on these intelligent mapping results, three to five verification routes are designed that cross major geological boundary lines to control the boundaries of regional geological bodies and optimize boundary identification accuracy. Based on the comprehensive geological data, geological boundaries are delineated according to stratigraphic occurrence and topography, using the V-shaped rule. This results in a comprehensive geological interpretation map that identifies geological unit boundaries (tectonic lines, classification unit boundaries, deformation patterns, etc.) and unique geological bodies (strata, rock masses, etc.).

[0086] As an embodiment, the present invention will be further described below by taking a 1:50,000 carp pond map of Hunan Province as an example.

[0087] The data for the Liyutang area in Hunan Province primarily includes basic geological data, geophysical data, geochemical data, remote sensing data, as well as publicly available papers and test and analysis data. The data used in this example primarily includes the total radioactive gamma spectrum collected by an airborne gamma spectrometer, aeromagnetic data collected by an airborne magnetometer, and basic geological maps.

[0088] Figure 3 It is a basic geological map, including 26 different geological structures such as the Holocene alluvial layer and the Paleogene Zaoshi Formation.

[0089] Figure 4 (a) is the total amount of radioactive gamma-ray spectrum data, denoted as γ, which can reflect the distribution characteristics of rock masses and strata with large radioactivity differences; (b) is the magnetotelluric data; and (c) is the gamma-ray spectrum and magnetotelluric interpretation line.

[0090] Figure 5The regional aeromagnetic anomalies and residual aeromagnetic anomalies that reveal the regional magnetic basement fluctuations and the distribution of deep tectonic units, as well as the reaction fault system, the distribution of igneous rocks and the fluctuations of magnetic strata shown in the (including data obtained by processing and converting magnetic anomalies in order to highlight certain information and suppress other information, such as ΔT post-pole magnetic anomalies, ΔT magnetic anomaly post-pole extension 2 km residual field, third-order details, etc.).

[0091] The training of the network consists of the following steps:

[0092] (1) Data preprocessing: Based on the basic geological map of Liyutang total MATLAB software was used to create a true value label matrix of 4338×5870. The total radioactive gamma ray spectrum data and aeromagnetic anomaly data were processed through magnetic polarization, analytical continuation, and wavelet multiscale analysis. These data were then converted into matrix format, resulting in a 4338×5870 total radioactive gamma ray spectrum matrix and a 4338×5870 aeromagnetic matrix, respectively. Experts interpreted the collected data. Finally, the data format and size were unified, enabling the implementation of supervised learning methods based on the U-Net network.

[0093] (2) The first stage - coarse classification stage: according to Ω total Divide similar geological types and obtain 5 mapping areas to be subdivided Ω k (k=1,2,...,5) and 5 categories of coarse classification labels S gt_stage1 ,like Figure 6 As shown in (a), Ω k (k=1,2,...,5) are the Paleogene area (Note: the bedrock of the Paleogene is the Carboniferous), Cretaceous area, Devonian-Carboniferous area, Sinian-Cambrian area, and granite area respectively.

[0094] (3) Select the total amount of radioactive gamma energy spectrum (denoted as γ) and the interpretation line data as the experimental data of the coarse stage, and fuse these data in the input layer to form a two-channel input. Figure 6 The area corresponding to the red box in (a) is used as training data, and the rest is test data. Using the random cropping method, the training data is divided into 10,000 128×128 size sub-blocks as the training data set S input_stage1 :{S1,S2,...,S 10000}, the ground-truth geological map corresponding to the training area is the predicted label S label_stage1 :{P1,P2,...,P 10000}.

[0095] (4) Through S input_stage1 and S label_stage1Training U-Net network F1, the loss function uses the cross entropy loss function, and the formula for network parameter optimization is as follows:

[0096]

[0097] Use the Adam algorithm to update the network parameters, minimize the loss, and then obtain the optimized network parameters The network was trained for 50 rounds. After each round of training, the performance of the current model was tested on the area to be filled. The overall accuracy (ACC) was used to evaluate the model performance. The formula is as follows:

[0098]

[0099] Among them, Num correct The number of correct pixels predicted by the network, Num all For S gt_stage1 The total number of pixels in the network is calculated, and the model with the highest accuracy is selected as the final network.

[0100] (5) In the input layer, γ and interpretation lines are fused to form two channels, which are input into the trained U-Net network to predict the five geological types of the carp pond area and obtain the intelligent mapping results of the coarse classification stage.

[0101] like Figure 6 As shown in (b), the prediction accuracy of the first stage 5-category coarse classification is 95.43%, and the large block contours and boundary features have been clearly segmented.

[0102] (6) The second stage - subdivision stage: Since the ΔT-type post-pole magnetic anomaly (denoted as MA), the ΔT magnetic anomaly post-pole extension 2 km residual field (denoted as RF), and the third-order details in the voyage anomaly wavelet multi-scale analysis (denoted as third-order details) can reflect the detailed characteristics of different geological types, they are combined with γ in different data combinations to form different input data combinations to train the U-Net network, realizing the Ω k (k=1,2,...,5) subdivisions. Among them, the Paleogene area contains 3 geological types, the Cretaceous area contains 6 geological types, the Devonian-Carboniferous area contains 7 geological types, the Sinian-Cambrian area contains 11 geological types, and the granite area contains 3 geological types. Select appropriate training areas respectively to cover every subdivision of geological types as much as possible. Figure 7As shown in (a), for the Paleogene region, geophysical data corresponding to the red box was selected as training data. Similarly, the areas corresponding to the black, purple, blue, and yellow boxes were selected as training data for the Cretaceous, Devonian-Carboniferous, Sinian-Cambrian, and granite, respectively. Separate networks 2, 3, ..., n+1 were then trained for each of these five coarsely classified regions. The optimal data combination and the optimal segmentation accuracy for each of the five coarsely classified regions are shown in Table 1.

[0103] Table 1 Best experimental results in the segmentation stage

[0104]

[0105] (7) Finally, the intelligent mapping results of the detailed classification stage are integrated into the intelligent mapping results of the coarse classification stage to obtain the final 26 categories of intelligent geological mapping results. Figure 7 As shown in (b), the accuracy rate is 80.08%. The accuracy of the Paleogene is 86.52%, and the accuracy of the Cretaceous is 85.34%. The outlines and boundaries are well visible. The accuracy of the Devonian-Carboniferous is 77.76%. The general outlines are reflected in the segmentation results, but due to the small amount of training data and the large number of geological types, the details are not clearly segmented. The accuracy of the Sinian-Cambrian region is 70.30%. Due to the large number of geological types and complex structure, the detailed layered information is not fully reflected. Further data that can reflect this information is needed. The accuracy of the granite region reaches 95.75%.

[0106] (8) Field verification. Verification is carried out through manually marked occurrence points. Figure 8 This is a prediction result diagram with occurrence point marks. It can be seen from the diagram that the prediction results of the occurrence points are basically accurate.

[0107] (9) Expert sketching. Geological experts, based on their professional knowledge, process the 26 types of intelligent geological mapping results through sketching to obtain the following: Figure 9 A geological interpretation map showing the boundaries of geological units and identification of special geological bodies is shown.

[0108] The two-stage multi-source data fusion method of the present invention uses the classic U-Net network as the basic network, fully considering that geophysical data such as aeromagnetic anomaly data can reflect rich characteristic information of different geological types, improving the accuracy of geological mapping under limited sample conditions, while saving manpower, material and financial resources, and having certain guiding significance for the mapping work of covered areas and restricted working areas, thereby improving the efficiency of geological mapping.

[0109] The beneficial effects of the present invention are as follows: the two-stage multi-source data fusion geological mapping method based on the U-Net network of the present invention utilizes the collected multi-source data and uses a deep learning method to automatically extract the characteristics of various data, fully explore the relationship between the data and different geological types, and adopts a two-stage strategy to complete intelligent geological mapping, thereby assisting geological experts in carrying out geological surveys.

[0110] The algorithm and implementation process proposed in this invention can utilize a smaller training area and accelerate the prediction of the area to be mapped in a GPU environment. Experimental results on the 1:50,000 Liyutang map of Hunan Province show that the algorithm has high resolution and generalization ability, and has certain guiding significance for the mapping work of covered areas and restricted working areas.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A two-stage multi-source data fusion geological mapping method based on U-Net network, characterized by: The following steps are involved: S1. Obtain multi-source geological data of the area to be mapped, and record the basic geological map of the area to be mapped as Ω total ; S2. preprocessing the multi-source geological data to obtain preprocessed multi-source geological data; S3, using U-Net neural network to map the basic geological map of the entire area to be mapped total Perform coarse classification to obtain the first stage of coarse classification intelligent mapping results; S4, using the U-Net neural network to subdivide the first stage's coarse intelligent mapping results to obtain the second stage's subdivided intelligent mapping results; S5. Integrate the second-stage subdivided intelligent mapping results into the first-stage coarse-divided intelligent mapping results to obtain the basic geological map Ω of the entire area to be mapped. total Intelligent mapping results; S6. Basic geological map of the entire area to be mapped Ω total The intelligent mapping results are verified in the field to obtain the revised intelligent mapping results; S7. Complete the geological interpretation map based on the corrected intelligent mapping results.

2. The two-stage multi-source data fusion geological mapping method based on the U-Net network according to claim 1, characterized in that: Step S3 is specifically as follows: S31: Get training set S input_stage1 , and the training set S input_stage1 The corresponding predicted label is recorded as S label_stage1 ; S32: Constructing a U-Net neural network for coarse geological type prediction; S33: Using training set S input_stage1 and the corresponding predicted label S label_stage1 , train the U-Net neural network for coarse geological type prediction, obtain the prediction loss, minimize the loss and obtain the optimized network parameters S34: Based on the optimized network parameters Obtain the trained U-Net neural network for coarse geological type prediction; S35: Input the pre-processed multi-source geological data of the entire area to be mapped into the trained coarse geological type prediction U-Net neural network to obtain the first stage coarse classification intelligent mapping result, so that the basic geological map of the entire area to be mapped Ω total Divided into n types of coarse-grained regions Ω k (k=1,2,...,n).

3. The two-stage multi-source data fusion geological mapping method based on the U-Net network according to claim 2, characterized in that: Step S4 is specifically as follows: S41: Get the training set and classify the regions into n types according to Ω k (k=1,2,...,n) Randomly cut the preprocessed multi-source geological data corresponding to the training set into multiple N×N sub-blocks to obtain n sub-training sets The corresponding predicted label is recorded as S43: Constructing n sub-segment geological type prediction U-Net neural network; S44: Using n sub-training sets, train n sub-segmented geological type prediction U-Net neural networks one by one, obtain prediction loss, and minimize the loss to obtain optimized network parameters S45: Based on the optimized network parameters Obtain the trained n sub-segment geological type prediction U-Net neural network; S46: n types of roughly divided regions Ω k The multi-source data corresponding to (k=1,2,...,n) are respectively input into the corresponding trained n sub-segment geological type prediction U-Net neural network to obtain the second stage subdivision intelligent mapping result Ω k1 ,Ω k2 ,..., 4. The two-stage multi-source data fusion geological mapping method based on the U-Net network according to claim 3, characterized in that: The optimized network parameters involved in step S33 and step S44 Its general expression formula is as follows: Among them, F is the coarse geological type prediction U-Net neural network or one of the n sub-segmented geological type prediction U-Net neural networks, S label For S label_stage1 or x∈Φ, is the data point in the training set; the mapping l:Φ→{1,2,...,M} represents the true category corresponding to each data point.

Images