Clustering-combined hyperspectral remote sensing deep learning lithology mapping method
By combining clustering and deep learning methods, lithologic mapping of hyperspectral remote sensing data is solved, and the problem of sample selection and efficiency improvement in large-scale lithologic mapping is achieved, achieving high-precision and efficient lithologic classification.
Patent Information
- Application Number
- CN202411554016.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-02-18
AI Technical Summary
In hyperspectral remote sensing lithologic maps, the prior art is difficult to effectively solve the problem of sample selection and efficiency improvement in large-scale lithologic maps.
Using hyperspectral remote sensing deep learning lithogram method combined with clustering, the hyperspectral data is preprocessed, lithogenetic clustering is carried out using the fuzzy C-mean method, regional lithogenetic sample library is established, and deep learning neural network model is combined with Lithological features were extracted and mapped.
This method effectively improves the efficiency of lithologic sample selection, ensures the accuracy of the sample, and achieves more refined and efficient lithologic classification through deep learning, which is in line with the actual surface lithologic distribution.
Smart Images
Figure CN119478127A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of hyperspectral remote sensing geology, and specifically relates to a hyperspectral remote sensing deep learning lithology mapping method combined with clustering. Background Art
[0002] With the continuous development of artificial intelligence technology, deep learning methods are increasingly applied to the field of remote sensing geology, and have become a research hotspot in lithology mapping. In terms of hyperspectral remote sensing target recognition based on deep learning, most of the research is carried out on hyperspectral remote sensing standard data sets, which have two obvious characteristics: one is that the standard data set has a small range, and the other is that there are known samples. Faced with the large-scale and medium-scale hyperspectral remote sensing lithology mapping needs, how to obtain accurate lithology samples is a key step in hyperspectral remote sensing lithology mapping.
[0003] Hyperspectral remote sensing lithology identification methods can be divided into two categories: unsupervised clustering methods and supervised classification methods with samples. The former can be classified without lithology samples, such as k-means, Isodata, FCM, DBSCAN, graph clustering, etc. Although no prior samples are required, the number of classifications needs to be specified in advance, and repeated experiments are required to find the optimal number of classifications; the latter requires the preparation of lithology samples in advance, such as neural network deep learning methods. Although the classification and identification accuracy is significantly higher than the former, it requires a large number of accurate samples to support it.
[0004] Therefore, there is an urgent need to develop a comprehensive technical method that fully utilizes the respective advantages of clustering and deep learning to solve the problems of sample selection and efficiency improvement faced by large-scale lithology mapping using hyperspectral remote sensing. Summary of the invention
[0005] The purpose of the present invention is to provide a hyperspectral remote sensing deep learning lithology mapping method combined with clustering, which can effectively solve the difficult problems of sample selection and efficiency improvement in large-scale hyperspectral remote sensing lithology mapping.
[0006] The technical solution to achieve the purpose of the present invention is:
[0007] A hyperspectral remote sensing deep learning lithology mapping method combined with clustering, the method comprising:
[0008] Step 1: Preprocess the hyperspectral data to obtain the target hyperspectral reflectance data;
[0009] Step 2: Calculate the number of lithology clusters, input the cluster number into the fuzzy C-means method, perform lithology clustering on the acquired hyperspectral reflectance data, and distinguish different types of lithology geological bodies;
[0010] Step 3: Determine the lithology of the different types of geological bodies that have been distinguished;
[0011] Step 4: Extract patches as samples for each lithology and establish a regional lithology sample library;
[0012] Step 5: Based on the acquired hyperspectral reflectance data and the established regional lithology sample library, the deep learning neural network model is trained to obtain the lithology feature extraction network model for lithology mapping.
[0013] The preprocessing of the hyperspectral data in step 1 includes radiation calibration, atmospheric correction and terrain correction.
[0014] The second step comprises:
[0015] Step 2.1: According to the number of lithology types in the geological map, set the initial value of the fuzzy factor and the initial number of clusters;
[0016] Step 2.2: Calculate cluster centers;
[0017] Step 2.3: Calculate the fuzzy segmentation entropy;
[0018] Step 2.4: Count the number of clusters in different cluster centers, and determine the best fuzzy factor based on the fuzzy segmentation entropy. The number of clusters corresponding to the best fuzzy factor is the initial lithology cluster number.
[0019] Step 2.5: Use the fuzzy C-means method to set the number of lithology clusters, perform lithology clustering on the acquired hyperspectral reflectance data, and obtain the lithology clustering results.
[0020] The calculation formula of the cluster center in step 2.2 is:
[0021]
[0022] Where k is the iteration number; j is the initial clustering category index, j = 1, 2…, c; is the cluster center of the jth category after iteration k; is the cluster center of the jth category at iteration k+1; x i is the spectral vector of pixel i in the hyperspectral reflectance data; i is the pixel index, i = 1, 2…, n; c is the number of clusters; m is the fuzzy factor; δ is the Euclidean distance; n is the number of pixels in the hyperspectral reflectance data.
[0023] The calculation formula of the fuzzy segmentation entropy in step 2.3 is:
[0024]
[0025] Where n is the number of pixels in the hyperspectral reflectance data; c is the number of clusters; u ij is the membership degree of the jth pixel in the image to the i-th category;
[0026] Membership degree u ij The calculation formula is:
[0027]
[0028] In the formula, c is the number of clusters; m is the fuzzy factor; δ is the Euclidean distance; v j is the jth cluster center; x i is the spectral vector of pixel i in the hyperspectral reflectance data.
[0029] The step 2.4 is specifically as follows: counting the number of different cluster center values to obtain the actual number of categories c under a specific m value, determining the optimal m based on the relationship between the fuzzy factor m and the fuzzy segmentation entropy PE, and the m value corresponding to the turning point when PE changes from a sharp drop to a slow drop is the optimal m. At this time, the number of clusters c corresponding to the optimal m is the initial lithology clustering number of the fuzzy C-means method.
[0030] The step three comprises:
[0031] Refer to the geological map to preliminarily determine the lithology types corresponding to the different geological body categories. If the known lithology range of the geological map does not match the clustering result, conduct field investigation and evidence collection to determine the specific lithology of the clustering category; or
[0032] The hyperspectral reflectance spectrum curve corresponding to each type of geological body is extracted from the clustering results, and the spectrum curve is matched with the known lithology spectrum, and the specific lithology category of the clustering results is determined according to the spectral matching rate.
[0033] The lithology sample library established in step 4 is usually a vector file in SHP format, with coordinate information and attributes including lithology categories.
[0034] The step five comprises:
[0035] Step 5.1: Create training set grids and prediction set grids based on hyperspectral reflectance data and lithology samples;
[0036] Step 5.2: Use the hyperspectral reflectance data and the training set raster data to train the deep learning neural network model to obtain the lithology feature extraction network model;
[0037] Step 5.3: Using the hyperspectral reflectance data and the prediction set raster data, the lithology feature extraction network model is used to perform lithology mapping and identification.
[0038] The beneficial technical effects of the present invention are:
[0039] 1. The present invention provides a hyperspectral remote sensing deep learning lithology mapping method combined with clustering. It uses the results of hyperspectral clustering analysis combined with field investigation verification to ensure the accuracy of lithology samples and improve the efficiency of lithology sample selection, providing a technical basis for using hyperspectral data to carry out large-area lithology deep learning mapping applications.
[0040] 2. The present invention provides a hyperspectral remote sensing deep learning lithology mapping method combined with clustering. The deep learning lithology classification combined with clustering is more refined and more efficient. It has been verified through field investigation and is found to be more in line with the actual surface lithology distribution. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A flow chart of a hyperspectral remote sensing deep learning lithology mapping method combined with clustering provided by the present invention;
[0042] Figure 2 A schematic diagram of the structure of the deep learning network model for lithology extraction provided by the present invention;
[0043] Figure 3 Hyperspectral lithology clustering and deep learning mapping provided by the present invention and geological map lithology comparison diagram: Figure 3 a is a geological map lithology division map, which provides a lithology reference for this embodiment. Figure 3 b is the fuzzy C-means clustering result, Figure 3 c is the lithology mapping result of the deep convolutional neural network model in this embodiment. DETAILED DESCRIPTION
[0044] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.
[0045] like Figure 1 As shown, the present invention provides a hyperspectral remote sensing deep learning lithology mapping method combined with clustering, which specifically includes the following steps:
[0046] Step 1: Preprocess the hyperspectral data to obtain the target hyperspectral reflectance data.
[0047] Preprocessing of hyperspectral data includes radiation calibration, atmospheric correction and terrain correction. The real reflectance value of the ground object is obtained through radiation calibration and atmospheric correction, and the influence of terrain undulation on lithology clustering is eliminated or reduced through terrain correction.
[0048] Step 2: Calculate the number of lithology clusters and input the cluster number into the fuzzy C-means method to perform lithology clustering on the hyperspectral reflectance data obtained in step 1 to distinguish different types of lithology geological bodies.
[0049] The fuzzy C-means method is used for lithology clustering. This method assigns fuzzy membership to each pixel, divides the pixels into different cluster centers, and allows pixels to belong to multiple clusters at the same time. This feature is more expressive for attributes with similar spectral characteristics between different lithologies. The key to lithology clustering using the fuzzy C-means method is to determine the number of clusters, which is determined by the following method:
[0050] Fuzzy segmentation entropy is used as the evaluation index, and the best fuzzy factor is selected. The number of clusters corresponding to the best fuzzy factor is the number of clusters.
[0051] The calculation formula of fuzzy segmentation entropy PE is as follows:
[0052]
[0053] In the above formula, n is the number of pixels of hyperspectral reflectance data, c is the number of clusters, and u is ij Indicates the membership degree of the j-th pixel in the image to the i-th category.
[0054] Membership degree u ij Calculated by the following formula:
[0055]
[0056] In the above formula, m is the fuzzy factor, δ is the Euclidean distance, and v j is the jth cluster center, x i is the spectral vector of pixel i in the hyperspectral reflectance data, i is the pixel index, i = 1, 2…, n, and c is the number of clusters.
[0057] Cluster center v j Calculated by the following formula:
[0058]
[0059] In the above formula, k is the iteration number, j is the initial clustering category index, j = 1, 2..., c, is the cluster center of the jth category after iteration k, is the cluster center of the jth category at iteration k+1.
[0060] when When it is less than a given threshold or iterates k times, the iteration stops.
[0061] Step 2.1: According to the number of lithology types in the geological map, set the initial value of the fuzzy factor and the initial number of clusters and other parameters
[0062] Referring to the lithology types of geological maps within the range of hyperspectral remote sensing data, half of the number of lithology types in the geological map is the initial value of the fuzzy factor m, and the number of increments is 2 to reduce the amount of calculation caused by the increment of m from 1. 1.5 times the number of lithology types in the geological map is the initial cluster number. The number of iterations k is set to 100, and the Euclidean distance δ is set to 1e-3.
[0063] Step 2.2: Calculate cluster centers
[0064] The cluster center of each category is calculated by formula (3). The initial value of the fuzzy factor m is half of the number of lithology in the geological map, the number of increments of m is 2, and the initial number of clusters is 1.5 times the number of lithology types in the geological map. The cluster centers of different categories are calculated iteratively by increasing the k value until When it is less than δ or iterates k times, stop iterating.
[0065] Step 2.3: Calculate fuzzy segmentation entropy
[0066] The fuzzy segmentation entropy PE is calculated by formula (1).
[0067] Step 2.4: Count the number of clusters with different cluster centers, and determine the optimal fuzzy factor based on the fuzzy segmentation entropy. The number of clusters corresponding to the optimal fuzzy factor is the lithology cluster number c.
[0068] The number of different cluster center values is counted to obtain the actual number of categories c under a specific m value. The optimal m is determined based on the relationship between the fuzzy factor m and the fuzzy segmentation entropy PE. The m value corresponding to the turning point when PE changes from a sharp decline to a slow decline is the optimal m. At this time, the number of clusters c corresponding to the optimal m is the lithology clustering number of the fuzzy C-means method.
[0069] Step 2.5: Use the fuzzy C-means method to set the number of lithology clusters, perform lithology clustering on the hyperspectral reflectance data obtained in step 1, and obtain the lithology clustering results.
[0070] Step 3: Determine the lithology types of the different types of geological bodies identified in step 2.
[0071] Two methods are provided here to determine the lithology types of different types of geological bodies distinguished in step two. The first method: refer to the geological map to preliminarily determine the lithology types corresponding to different geological body categories in the clustering results of step two. In the case where the known lithology range of the geological map does not match the clustering results, conduct field investigations and evidence collection to determine the specific lithology of the clustering categories. The lithology information of different clustering categories is determined by combining reference to geological maps and field investigations, thereby ensuring the accuracy of the lithology categories. The second method: extract the high-spectral reflectance spectral curve corresponding to each type of geological body from the clustering results of step two, match the spectral curve with the known lithology spectrum, and determine the specific lithology category of the clustering results based on the spectral matching rate. This method requires a certain working foundation and the collection of spectral curves of different lithology types in advance.
[0072] Due to the influence of weathering and alteration, the same type of lithology often clusters into different categories. In addition, the shadow of steeper mountains will also cluster different types of lithology. For the former, the method of further subdividing the altered lithology is used, or the same type of lithology is merged; for the latter, the same type of lithology is merged. This optimization process lays a good foundation for the subsequent selection of lithology samples.
[0073] Step 4: Extract patches as samples for each lithology and establish a regional lithology sample library.
[0074] On the basis of determining the clustered lithology types in step 3, a certain number of patches are extracted for each lithology as samples to establish a regional lithology sample library.
[0075] When delineating lithology, try to avoid valleys, because valleys often contain different types of rocks and usually have a certain amount of vegetation. For non-lithology targets such as larger areas of vegetation and water bodies, separate samples should be established.
[0076] The lithology sample library established by this method is usually a vector file in SHP format, with coordinate information and lithology categories in its attributes. Extracting lithology samples by the above method not only reduces the workload of field investigation and verification, improves work efficiency, but also ensures the accuracy of lithology samples.
[0077] Step 5: Based on the acquired hyperspectral reflectance data and the establishment of a regional lithology sample library, the deep learning neural network model is trained to obtain a lithology feature extraction network model for lithology mapping.
[0078] The hyperspectral reflectance data processed in step one and the lithology samples established in step four are used as input to train the deep learning neural network model, obtain the lithology feature extraction network model, and perform lithology mapping.
[0079] Step 5.1: Create training set grids and prediction set grids based on hyperspectral reflectance data and lithology samples
[0080] Before deep learning model training, it is necessary to produce training set rasters and prediction set rasters based on hyperspectral reflectance data and lithology samples.
[0081] First, assign a corresponding lithology code to each type of lithology, and the lithology codes are numbered starting from 1.
[0082] Furthermore, the lithology samples are converted into training set grids according to the range and resolution of the hyperspectral data, and the non-sample areas in the training set grids are assigned a value of 0.
[0083] Furthermore, the 0 value in the training set grid is converted to a value greater than the number of lithology types (such as 100), and then the hyperspectral data is used as a mask for clipping to generate the prediction set grid.
[0084] Step 5.2: Use hyperspectral reflectance data and training set raster data to train the deep learning neural network model to obtain the lithology feature extraction network model
[0085] according to Figure 2 The structure shown in the figure and the parameters shown in Table 1 are used to construct a deep learning neural network model for lithology feature extraction, which is based on the pytorch deep learning framework. The processed hyperspectral reflectance data and the training set raster data prepared by the above method are used as input, and the hyperspectral reflectance data are sliced into 3×3×N size to form multiple spectral data cubes as training data. The three-dimensional convolutional neural network is trained to obtain the lithology feature extraction network model.
[0086] Table 1. Lithology extraction deep learning network model architecture and parameters
[0087]
[0088] The three-dimensional convolutional neural network in this embodiment includes two convolutional layers, namely C1 and C2, and the convolution kernels are all three-dimensional convolution kernels. The first three-dimensional convolutional layer C1 includes two three-dimensional kernels and outputs two three-dimensional feature vectors as the input of the second layer; the second three-dimensional convolutional layer C2 includes four three-dimensional kernels and generates eight three-dimensional feature vectors; the output of the C2 layer is flattened into a feature vector and input to the fully connected layer F1, where the output feature vector contains the final learned spectral-spatial features. This deep learning method has no pooling layer, does not perform spatial feature dimensionality reduction, retains pixel-level spatial-spectral features, has the advantages of few parameters, light model and strong ability to avoid overfitting, and is easier to train hyperspectral big data.
[0089] During the model training process, the network model divides the training set into a training set and a validation set in a ratio of 9:1. During the model training process, the training round number epoch and the convergence of the loss function (i.e., the validation set classification accuracy) are used to determine whether the training is completed. Model training ends when one of the training rounds and the loss function meets the set requirements. If the number of training rounds is set too small, the network model does not converge, and the sample feature training is insufficient; if the number of training rounds is set too large, it will waste computing resources and there will be a risk of overfitting. When the loss function tends to be stable and no longer fluctuates with the increase in the number of training rounds, the number of training rounds at this time is optimal, and the number of training rounds in this embodiment is set to 150. In addition, the batch_size and learning rate learning_rate of each training sample are also important parameters of the model. Batch_size affects the convergence of the model, and learning_rate affects the convergence speed of the model. Improper settings cannot achieve local optimality. In this embodiment, the number of training samples per training is set to 128, and the learning rate is set to 0.01.
[0090] Step 5.3: Using hyperspectral reflectance data and prediction set raster data, and using lithology feature extraction network model, perform lithology mapping and identification
[0091] The processed hyperspectral reflectance data and prediction set raster data are further input as prediction targets, the trained lithology extraction network model is used for lithology mapping and identification, and the validation set is used to evaluate the accuracy of the classification results.
[0092] Figure 3 Shown is a comparison between the classification results of the deep learning network model for hyperspectral lithology feature extraction and the geological map. Figure 3 a is a geological map lithology division map, which provides a lithology reference for this embodiment. Figure 3 b is the fuzzy C-means clustering result, Figure 3 c is the lithology mapping result of the deep convolutional neural network model in this embodiment.
[0093] Figure 3 The results show that the lithology clustering results provided by the present invention provide good lithology samples for deep learning and improve work efficiency; the deep learning lithology classification combined with clustering provided by the present invention is more refined and more efficient, and it has been verified by field investigation that it is more in line with the actual surface lithology distribution.
[0094] The present invention is described in detail above with reference to the accompanying drawings and embodiments, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of ordinary technicians in the field without departing from the purpose of the present invention. The contents not described in detail in the present invention can adopt the existing technology.
Claims
1. A hyperspectral remote sensing deep learning lithology mapping method combined with clustering, characterized in that: The method comprises: Step 1: Preprocess the hyperspectral data to obtain the target hyperspectral reflectance data; Step 2: Calculate the number of lithology clusters, input the cluster number into the fuzzy C-means method, perform lithology clustering on the acquired hyperspectral reflectance data, and distinguish different types of lithology geological bodies; Step 3: Determine the lithology of the different types of geological bodies that have been distinguished; Step 4: Extract patches as samples for each lithology and establish a regional lithology sample library; Step 5: Based on the acquired hyperspectral reflectance data and the established regional lithology sample library, the deep learning neural network model is trained to obtain the lithology feature extraction network model for lithology mapping.
2. The method for deep learning lithology mapping based on hyperspectral remote sensing combined with clustering according to claim 1, characterized in that: The preprocessing of the hyperspectral data in step 1 includes radiation calibration, atmospheric correction and terrain correction.
3. The method for deep learning lithology mapping based on hyperspectral remote sensing combined with clustering according to claim 1, characterized in that: The second step comprises: Step 2.1: According to the number of lithology types in the geological map, set the initial value of the fuzzy factor and the initial number of clusters; Step 2.2: Calculate cluster centers; Step 2.3: Calculate the fuzzy segmentation entropy; Step 2.4: Count the number of clusters in different cluster centers, and determine the best fuzzy factor based on the fuzzy segmentation entropy. The number of clusters corresponding to the best fuzzy factor is the initial lithology cluster number. Step 2.5: Use the fuzzy C-means method to set the number of lithology clusters, perform lithology clustering on the acquired hyperspectral reflectance data, and obtain the lithology clustering results.
4. The method for deep learning lithology mapping based on hyperspectral remote sensing combined with clustering according to claim 3, characterized in that: The calculation formula of the cluster center in step 2.2 is: Where k is the iteration number; j is the initial clustering category index, j = 1, 2…, c; is the cluster center of the jth category after iteration k; is the cluster center of the jth category at iteration k+1; x i is the spectral vector of pixel i in the hyperspectral reflectance data; i is the pixel index, i = 1, 2…, n; c is the number of clusters; m is the fuzzy factor; δ is the Euclidean distance; n is the number of pixels in the hyperspectral reflectance data.
5. The method for deep learning lithology mapping using hyperspectral remote sensing combined with clustering according to claim 4, characterized in that: The calculation formula of the fuzzy segmentation entropy in step 2.3 is: Where n is the number of pixels in the hyperspectral reflectance data; c is the number of clusters; u ij is the membership degree of the jth pixel in the image to the i-th category; Membership degree u ij The calculation formula is: In the formula, c is the number of clusters; m is the fuzzy factor; δ is the Euclidean distance; v j is the jth cluster center; x i is the spectral vector of pixel i in the hyperspectral reflectance data.
6. The method for deep learning lithology mapping using hyperspectral remote sensing combined with clustering according to claim 5, characterized in that: The step 2.4 is specifically as follows: counting the number of different cluster center values to obtain the actual number of categories c under a specific m value, determining the optimal m based on the relationship between the fuzzy factor m and the fuzzy segmentation entropy PE, and the m value corresponding to the turning point when PE changes from a sharp drop to a slow drop is the optimal m. At this time, the number of clusters c corresponding to the optimal m is the initial lithology clustering number of the fuzzy C-means method.
7. The method for deep learning lithology mapping using hyperspectral remote sensing combined with clustering according to claim 1, characterized in that: The step three comprises: Refer to the geological map to preliminarily determine the lithology types corresponding to the different geological body categories. If the known lithology range of the geological map does not match the clustering result, conduct field investigation and evidence collection to determine the specific lithology of the clustering category; or The hyperspectral reflectance spectrum curve corresponding to each type of geological body is extracted from the clustering results, and the spectrum curve is matched with the known lithology spectrum, and the specific lithology category of the clustering results is determined according to the spectral matching rate.
8. The method for deep learning lithology mapping using hyperspectral remote sensing combined with clustering according to claim 1, characterized in that: The lithology sample library established in step 4 is usually a vector file in SHP format, with coordinate information and attributes including lithology categories.
9. The method for deep learning lithology mapping using hyperspectral remote sensing combined with clustering according to claim 1, characterized in that: The step five comprises: Step 5.1: Create training set grids and prediction set grids based on hyperspectral reflectance data and lithology samples; Step 5.2: Use the hyperspectral reflectance data and the training set raster data to train the deep learning neural network model to obtain the lithology feature extraction network model; Step 5.3: Using the hyperspectral reflectance data and the prediction set raster data, the lithology feature extraction network model is used to perform lithology mapping and identification.