Multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network

Abstract

The invention belongs to the technical field of remote sensing image recognition, and discloses a multi-source remote sensing spodumene recognition method based on a spectral index and a double-branch neural network. The method comprises the steps of field sample collection, remote sensing image acquisition and pretreatment; calculating a spodumene index based on spectral characteristics of spodumene in a remote sensing image, and making a remote sensing index cube; a double-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network is constructed, spodumene spectral features and spatial distribution features are recognized, spodumene is further recognized, and recognition result precision evaluation is performed. According to the method, the problem that the spectral characteristics of spodumene in visible light-near infrared and short wave infrared band remote sensing images are not obvious and are difficult to identify is solved, a mixed deep learning model, namely a double-branch neural network model is used for fully utilizing graph and spectrum information in the remote sensing images, and the spodumene identification accuracy and efficiency are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of image remote sensing, and in particular relates to a multi-source remote sensing spodumene identification method based on spectral index and a double-branch neural network. Background Art

[0002] Lithium (Li) is the lightest metal in nature, with high specific heat, electrical conductivity and electrochemical activity. As an important component of green energy storage technology, it has important economic significance. Lithium has been listed as a strategic or key mineral by many countries and has been identified as a "green rare metal" by the United Nations Environment Programme. Global lithium resources are mainly found in salt lake brine (72.3%), and lithium resources are less in ores (20.3%), but lithium-containing minerals contribute more than 50% of the global lithium supply. Among the granite pegmatite-type lithium deposits, spodumene is the most important of the several key lithium-containing minerals where lithium is mainly found. Rapid identification of spodumene and strengthening the development and utilization of lithium resources are crucial to ensuring the stability of lithium supply and the sustainable development of global lithium resources. Remote sensing technology has the ability to accurately identify and classify and the potential to discover prospecting targets in complex geological environments. It has been widely used in rock and mineral identification and classification in the geological field, and has achieved remarkable results. Currently, remote sensing technology using visible light-shortwave infrared bands has made significant progress in the detection of various minerals, but minerals such as spodumene are anhydrous silicate minerals, and their spectral characteristics are mainly concentrated in the thermal infrared band. However, the application of lithium ore detection based on thermal infrared remote sensing is still in its early stages. How to accurately identify spodumene with thermal infrared remote sensing, and then distinguish between mineral-bearing and mineral-free pegmatites, faces challenges.

[0003] Through the above analysis, the problems and defects of the existing technology are: the spectral characteristics of spodumene are not obvious in the remote sensing images of visible light-near infrared and short-wave infrared bands, which makes it difficult to identify it with the existing technology, resulting in low accuracy and efficiency in spodumene identification. Summary of the invention

[0004] In order to overcome the problems existing in the related art, the disclosed embodiment of the present invention provides a multi-source remote sensing spodumene identification method integrating visible light, short-wave infrared and thermal infrared, and specifically relates to a multi-source remote sensing spodumene identification method based on spectral index and dual-branch neural network. The purpose of the present invention is to establish a series of spodumene remote sensing indexes based on multi-source satellite image data, realize accurate identification of spodumene through a hybrid deep learning model, and provide a new method for improving the efficiency of lithium resource detection.

[0005] The technical solution is as follows: A multi-source remote sensing spodumene identification method based on spectral index and dual-branch neural network includes:

[0006] S1, field sample collection, remote sensing image acquisition and preprocessing;

[0007] S2, based on the spectral characteristics of remote sensing images, calculate the spodumene spectral index and make a spodumene spectral remote sensing index cube;

[0008] S3, constructing a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network, identifying the spatial distribution characteristics and spectral characteristics of spodumene based on the two-branch neural network model, and obtaining an identification result graph;

[0009] S4, recognition result evaluation.

[0010] In step S1, field samples are collected, remote sensing images are acquired and preprocessed, including: for Landsat8 multispectral and GF-5 hyperspectral images, radiometric calibration and atmospheric correction are performed to convert the original values ​​into reflectance; for SDGSAT-1 thermal infrared images, radiometric calibration is performed to convert the original values ​​into radiometric brightness; after completion, the three images are aligned and resampled separately.

[0011] In step S2, the spodumene spectrum index is calculated and the spodumene spectrum remote sensing index cube is prepared, including:

[0012] S201, selecting the green band and the near infrared band in the visible light-near infrared range of the preprocessed Landsat 8 image to perform band ratio calculation to obtain the spodumene visible light-near infrared remote sensing index BR;

[0013] S202, selecting the short-wave infrared band of the pre-processed GF-5 image, extracting the spectrum of the pixel where spodumene is located in the collected sample and averaging it as the reference spectrum of the short-wave infrared band of spodumene, and using the matched filter method to calculate the matched filter value MF between all pixels and the reference spectrum as the short-wave infrared remote sensing index of spodumene;

[0014] S203, for the pre-processed SDGSAT-1 image, extract the three-band radiation brightness values ​​of the pixels where all spodumene samples are located, combine them two by two, and use the least squares method to fit the linear relationship of the thermal infrared radiation brightness of spodumene in each band combination;

[0015] S204, combining the visible light-near infrared remote sensing index, the short-wave infrared remote sensing index and multiple thermal infrared remote sensing indexes corresponding to each pixel to obtain a three-dimensional spodumene remote sensing index cube, which is arranged in pixel order and stored in a two-dimensional csv format.

[0016] In step S201, the calculation formula of spodumene visible light-near infrared remote sensing index BR is:

[0017]

[0018] In the formula, Band 3 is the green band and Band 5 is the near-infrared band.

[0019] In step S202, the calculation formula of the matched filter value MF is:

[0020]

[0021] In the formula, S i is the actual spectral value of the pixel in the i-th band, R i is the reference spectrum value of spodumene in the i-th band, and n is the total number of bands.

[0022] In step S202, the spodumene short-wave infrared remote sensing index ST i The calculation formula is:

[0023] ST i =|Band b -Band a ×m i -c i |

[0024] Where, Band a ,Band b are the two bands in the i-th band combination, m i ,c i They are the slope and intercept of the linear relationship obtained by linear fitting of spodumene in the th band combination.

[0025] In step S3, a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network is constructed, and the spatial distribution characteristics and spectral characteristics of spodumene are identified based on the two-branch neural network model, including:

[0026] S301, use Matlab software to produce the data required for training;

[0027] S302, build a deep learning model using Python language.

[0028] In step S301, Matlab software is used to prepare the data required for training, including:

[0029] (1) Read the image data and divide it into training set and test set according to the label of each pixel;

[0030] (2) Perform data conversion to convert the three-dimensional remote sensing image data and label data into two-dimensional matrices respectively;

[0031] (3) Calculate the Laplacian matrix required for graph convolutional networks;

[0032] (4) Save the training and test data as mat files respectively.

[0033] In step S302, a deep learning model is constructed using Python language, including:

[0034] (a) Import the Python libraries of Numpy, Tensorflow, and Scipy;

[0035] (b) Define auxiliary functions, including:

[0036] sample_mask, creates a sample mask for data segmentation;

[0037] create_placeholders, creates TensorFlow placeholders for input data and labels;

[0038] initialize_parameters, initialize network parameters, including weights and biases;

[0039] GCN_layer, defines the operation of the graph convolution layer, with the input being the feature matrix and the Laplacian matrix;

[0040] Mynetwork, defines the forward propagation process of the hybrid model, which contains GCN layers and 1D-CNN;

[0041] (c) Define the network structure, which consists of a graph convolution layer and a 1D-CNN. The graph convolution layer performs convolution operations on features through the Laplacian matrix, and the 1D-CNN is used to process the temporal information of the feature vector. Finally, the fully connected layer performs classification and outputs the probability distribution of the spodumene category.

[0042] (d) Use cross entropy as the loss function, add L2 regularization, and use Adam optimizer for parameter update;

[0043] (e) Use the leave-one-out cross-validation method to train the deep learning model, leaving one sample as the validation set each time and the rest as the training set. Then, train the deep learning model on the entire training set and make predictions on the test set to complete the spodumene identification.

[0044] (f) Save the predicted probability of the test set as a MAT file and a CSV file for subsequent analysis.

[0045] Another object of the present invention is to provide a multi-source remote sensing spodumene identification system based on spectral index and two-branch neural network, the system implements the multi-source remote sensing spodumene identification method based on spectral index and two-branch neural network, the system comprises:

[0046] Remote sensing image processing module, used for field sample collection, remote sensing image acquisition and preprocessing;

[0047] The remote sensing index cube production module is used to calculate the spodumene spectrum index based on the spectral characteristics of remote sensing images and produce the spodumene spectrum remote sensing index cube;

[0048] A two-branch neural network model construction module is used to construct a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network, and to identify the spatial distribution characteristics and spectral characteristics of spodumene based on the two-branch neural network model to obtain an identification result graph;

[0049] Evaluation module, used for identification result evaluation.

[0050] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows: the present invention solves the problem that spodumene is difficult to identify due to its unclear spectral characteristics in remote sensing images in the visible light-shortwave infrared band, and uses a hybrid deep learning model to fully utilize the image and spectrum information in the remote sensing images to improve the accuracy and efficiency of spodumene recognition.

[0051] Since the identification of spodumene is of great significance to the discovery of pegmatite-type lithium deposits, it is a direct sign of prospecting for pegmatite-type lithium deposits. Therefore, the present invention can realize the direct identification of the location and range of spodumene after the application and transformation of the results of the identification technology and method of spodumene, especially for high-altitude areas that are difficult to reach by manpower, it can greatly narrow the scope of prospecting and clarify the purpose of prospecting methods. For prospecting companies, it can not only save a lot of manpower and material resources and other prospecting costs, but also clarify the prospecting direction, promote the discovery of new mineral deposits, and generate economic benefits.

[0052] The present invention innovatively proposes spodumene spectral indices in the near-infrared, short-wave infrared and thermal infrared spectral bands, establishes a spodumene spectral index cube, and introduces a dual-branch neural network for the first time to identify spodumene, filling the technical gap of using thermal infrared remote sensing combined with other remote sensing data and deep learning algorithms to mine spodumene characteristics and then identify spodumene.

[0053] The short-wave infrared band mainly reflects the vibration between the hydroxyl group and the cation connected to it. As a chain anhydrous silicate mineral, spodumene has no absorption characteristics in the short-wave infrared band because there is no structural water in the crystal structure of spodumene. Since sericite is secondary distributed in the joints and fissures of spodumene, the spectral characteristics of spodumene in the short-wave infrared spectrum range are caused by mica-type hydrous minerals. Therefore, most researchers indirectly identify spodumene by identifying the spectral characteristics of spodumene. However, the diagnostic characteristics of spodumene are actually in the thermal infrared band, but there are few studies, and only some scholars have conducted research on the ground thermal infrared spectral characteristics of spodumene. Therefore, the present invention solves the technical problem that the spectral characteristics of spodumene are difficult to identify in the visible light-near infrared and short-wave infrared. By combining the thermal infrared and visible light-near infrared and short-wave infrared spectral bands, not only a spectral index reflecting the characteristics of spodumene is constructed, but also spodumene can be directly identified on remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0055] Figure 1 It is a multi-source remote sensing spodumene identification method and system diagram provided by an embodiment of the present invention;

[0056] Figure 2 It is a linear relationship diagram of thermal infrared radiation brightness in Band 1 and Band combination;

[0057] Figure 3 It is a linear relationship diagram of the thermal infrared radiation brightness in the combination of Band 1 and Band 3;

[0058] Figure 4 It is a linear relationship diagram of the thermal infrared radiation brightness in the combination of Band 2 and Band 3;

[0059] Figure 5 It is the Landsat 8BR map of spodumene remote sensing index raster image;

[0060] Figure 6 It is the GF-5MF map in the spodumene remote sensing index raster image;

[0061] Figure 7 It is the SDGSAT-1ST1 map in the spodumene remote sensing index raster image;

[0062] Figure 8 It is the SDGSAT-1ST2 map in the spodumene remote sensing index raster image;

[0063] Fig. 9It is the SDGSAT-1ST3 map in the spodumene remote sensing index raster image;

[0064] Fig.10 It is a diagram of the calculation process of a two-branch neural network;

[0065] Fig.11 This is the result of spodumene remote sensing identification. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0067] The innovation of the present invention lies in: creating a spodumene thermal infrared spectrum index for the first time, and establishing a two-branch neural network recognition model in the form of a spodumene light index cube of visible light-near infrared, short-wave infrared and thermal infrared for the first time, innovatively identifying the dual characteristics of spodumene spectral characteristics and spatial distribution characteristics, and significantly improving the recognition accuracy and efficiency of spodumene.

[0068] Embodiment 1, as Figure 1 As shown, the multi-source remote sensing spodumene identification method provided by the embodiment of the present invention includes:

[0069] S1, field sample collection, remote sensing image acquisition and preprocessing;

[0070] For example, spodumene and non-spodumene samples were uniformly collected in the vertical pegmatite veins in the study area; Landsat 8 multispectral, GF-5 hyperspectral and SDGSAT-1 thermal infrared remote sensing images of the study area were obtained respectively, and images of the same season, less cloud cover and better imaging quality were selected as much as possible; remote sensing image preprocessing was completed using ENVI 5.6 software, and for Landsat 8 multispectral and GF-5 hyperspectral images, radiometric calibration and atmospheric correction were performed to convert the original values ​​into reflectance, and for SDGSAT-1 thermal infrared images, radiometric calibration was performed to convert the original values ​​into radiometric brightness. After completion, the three images were registered and resampled to 30m×30m respectively.

[0071] S2, based on the spectral characteristics of remote sensing images, calculate the spodumene spectral index and make a spodumene spectral remote sensing index cube;

[0072] S201, select the green band (Band 3) and near infrared band (Band 5) of the visible light-near infrared range of the pre-processed Landsat 8 image to perform band ratio analysis, and obtain the spodumene visible light-near infrared remote sensing index BR, see Figure 5 Landsat 8BR map of spodumene remote sensing index raster image.

[0073]

[0074] In the formula, Band 3 is the green band and Band 5 is the near-infrared band.

[0075] S202, select the short-wave infrared band (2032-2496nm) of the pre-processed GF-5 image, extract the spectrum of the pixel where spodumene is located in the collected sample, take the average as the reference spectrum of the short-wave infrared band of spodumene, use the matched filtering method (Matched Filtering) to calculate the matched filtering value MF between all pixels and the reference spectrum, as the spodumene short-wave infrared remote sensing index, see Figure 6 GF-5MF map in the spodumene remote sensing index raster image.

[0076]

[0077] In the formula, S i is the actual spectral value of the pixel in the i-th band, R i is the reference spectrum value of spodumene in the i-th band, and n is the total number of bands.

[0078] S203, for the pre-processed SDGSAT-1 image, extract the three-band radiation brightness values ​​of the pixels where all spodumene samples are located, combine them two by two, and use the least squares method to fit the linear relationship of the thermal infrared radiation brightness of spodumene in each band combination (see Figure 2 The linear relationship between the thermal infrared radiation brightness of Band 1 and the combination of Bands in the SDGSAT-1 satellite thermal infrared band scatter plot and the spodumene fitting line is shown in Figure 2.

[0079] Figure 3 is the linear relationship between the thermal infrared radiation brightness in the combination of Band 1 and Band 3.

[0080] Figure 4 The linear relationship diagram of thermal infrared radiation brightness in the combination of Band 2 and Band 3 is used to calculate the spodumene thermal infrared remote sensing index ST i ,See Figure 7 SDGSAT-1ST1 map in the spodumene remote sensing index raster image, Figure 8 SDGSAT-1ST2 map in the spodumene remote sensing index raster image, Fig. 9 SDGSAT-1ST3 map of spodumene remote sensing index raster image.

[0081] ST i =|Band b -Band a ×m i -c i |

[0082] Where, Band a ,Band b are the two bands in the i-th band combination, m i ,c i They are the slope and intercept of the linear relationship obtained by linear fitting of spodumene in the th band combination.

[0083] S204, combining one visible light-near infrared remote sensing index, one short-wave infrared remote sensing index and three thermal infrared remote sensing indexes corresponding to each pixel to obtain a three-dimensional spodumene remote sensing index cube, which is arranged in pixel order and stored in a two-dimensional csv format.

[0084] S3, constructing a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network, identifying the spatial distribution characteristics and spectral characteristics of spodumene based on the two-branch neural network model, and obtaining an identification result graph;

[0085] Exemplarily, a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network is constructed, the spatial information of the spodumene remote sensing spectral index cube is extracted using the GCN model, and the spectral information of the spodumene remote sensing spectral index cube is extracted using 1D-CNN, and finally the spodumene recognition based on the two-branch neural network model is completed. The specific steps are as follows:

[0086] S301, use Matlab software to create the data required for training, including:

[0087] (1) Read the image data and divide it into training set and test set according to the label of each pixel;

[0088] (2) Perform data conversion, converting the three-dimensional remote sensing image data and label data into two-dimensional matrices respectively, and expanding all band information of each pixel into a vector so that each pixel occupies a row in the matrix. This conversion combines the spatial structure information and spectral information of the original image, facilitating subsequent feature processing and analysis;

[0089] (3) Calculate the Laplacian matrix required for the graph convolutional network. The Laplacian matrix is ​​constructed based on the similarity between samples and reflects the connection relationship between data points in the graph structure. By calculating the Laplacian matrix, the local geometric structure in the data can be captured, which is used to smooth and aggregate node features in the graph convolutional network, thereby improving the learning ability of the model;

[0090] (4) Save the training and test data as mat files respectively.

[0091] S302, using Python language to construct a recognition model of a two-branch neural network, including:

[0092] (a) Import necessary Python libraries such as Numpy, Tensorflow, and Scipy;

[0093] (b) Define auxiliary functions, including:

[0094] sample_mask: creates a sample mask for data segmentation;

[0095] create_placeholders: Create TensorFlow placeholders for input data and labels.

[0096] initialize_parameters: Initialize network parameters, including weights and biases;

[0097] GCN_layer: defines the operation of the graph convolution layer. The input is the feature matrix and the Laplacian matrix. The feature matrix is ​​first multiplied by the weight matrix and then by the Laplacian matrix to achieve feature smoothing and aggregation, thereby extracting local features in the graph structure.

[0098] mynetwork: defines the forward propagation process of the hybrid model. The input features are first convolved and smoothed through the graph convolution layer (GCN layer), and then extracted together with another part of the data through a one-dimensional convolutional neural network (1D-CNN);

[0099] (c) Define the network structure, which consists of a graph convolution layer and a 1D-CNN. The graph convolution layer performs convolution and smoothing operations on the features through the Laplacian matrix to extract local features in the graph structure. Subsequently, the 1D-CNN further convolves these features to extract higher-level spatial features, and then processes them through pooling and normalization. Finally, the fully connected layer performs classification and outputs the probability distribution of the spodumene category, that is, each grid corresponds to a probability value that the grid is spodumene. Through this probability distribution, the user can directly obtain the classification confidence of each input grid from the model output (see the specific process). Fig.10 Dual-branch neural network calculation process diagram);

[0100] (d) Use cross entropy as the loss function to measure the difference between the model prediction result and the true label, and introduce L2 regularization to prevent overfitting. The complexity of the model is limited by adding the square sum of weight parameters to the loss function. Finally, the Adam optimizer is used to optimize the total loss (the sum of cross entropy loss and L2 regularization loss). The loss is gradually reduced by adjusting the model parameters, thereby improving the accuracy and generalization ability of the model.

[0101] (e) The deep learning model is trained using the leave-one-out cross-validation method, that is, one sample is used as the validation set each time, and the remaining samples are used as the training set, and the model parameters are optimized round by round. In each round, the model updates the weights through forward propagation and back propagation, minimizes the loss function containing cross entropy loss and L2 regularization, and evaluates its performance on the validation set. After completing all cross-validation rounds, the model is finally trained on the entire training set to fully utilize all data to optimize the parameters. Subsequently, the model makes predictions on the test set, calculates the category probability distribution of each sample through forward propagation, and classifies and identifies the samples based on these probabilities, ultimately achieving accurate identification of spodumene. (See Fig.11 Spodumene remote sensing identification results); the overall parameters of the model are shown in Table 1.

[0102] Table 1 Core parameters of the dual-branch neural network model

[0103]

[0104]

[0105] (f) Save the predicted probability of the test set as a MAT file and a CSV file for subsequent analysis.

[0106] S4, Recognition result evaluation

[0107] For example, the model performance is evaluated, and indicators such as accuracy, precision, recall, and F1 score are calculated, and the AUC-ROC curve is plotted.

[0108] Accuracy measures the proportion of correctly classified samples among all samples. It takes into account the correctness of all predictions.

[0109]

[0110] Precision is the ratio of true positives to all predicted positives, indicating the model's ability to avoid false positives. It reflects the model's ability to avoid false alarms.

[0111]

[0112] Recall is the ratio of true positives to all actual positives and reflects the model's ability to identify relevant instances.

[0113] It reflects the model's ability to identify relevant instances.

[0114]

[0115] The F1 score is the harmonic mean of precision and recall, which provides a trade-off between precision and recall.

[0116]

[0117] AUC-ROC measures the model's ability to distinguish categories by evaluating the trade-off between the true positive rate (TPR) and the false positive rate (FPR) under different threshold settings.

[0118]

[0119] Among them, true positives (TP) are instances correctly identified as positive, false positives (FP) are instances incorrectly identified as positive, true negatives (TN) are instances correctly identified as negative, and false negatives (FN) are instances incorrectly identified as negative.

[0120] Compared with the use of visible light-near infrared index, short-wave infrared index and the combination of the two alone, the results of multi-source remote sensing index identification using spodumene thermal infrared remote sensing index achieved better results in all evaluation indicators, see Table 2.

[0121] Table 2 Quantitative evaluation results of models with different remote sensing index combinations

[0122]

[0123] Embodiment 2, the embodiment of the present invention provides a multi-source remote sensing spodumene identification system, the system comprising:

[0124] Remote sensing image processing module, used for field sample collection, remote sensing image acquisition and preprocessing;

[0125] The remote sensing index cube production module is used to calculate the spodumene spectrum index based on the spectral characteristics of remote sensing images and produce the spodumene spectrum remote sensing index cube;

[0126] A two-branch neural network model construction module is used to construct a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network, and to identify the spatial distribution characteristics and spectral characteristics of spodumene based on the two-branch neural network model to obtain an identification result graph;

[0127] Evaluation module, used for identification result evaluation.

[0128] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-source remote sensing spodumene identification method based on spectral index and dual-branch neural network, characterized in that: The method includes: S1, field sample collection, remote sensing image acquisition and preprocessing; S2, based on the spectral characteristics of remote sensing images, calculate the spodumene spectral index and make a spodumene spectral remote sensing index cube; S3, constructing a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network, identifying the spatial distribution characteristics and spectral characteristics of spodumene based on the two-branch neural network model, and obtaining an identification result graph; S4, recognition result evaluation.

2. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 1, characterized in that: In step S1, field samples are collected, remote sensing images are acquired and preprocessed, including: for Landsat 8 multispectral and GF-5 hyperspectral images, radiometric calibration and atmospheric correction are performed to convert the original values ​​into reflectance; for SDGSAT-1 thermal infrared images, radiometric calibration is performed to convert the original values ​​into radiometric brightness; after completion, the three images are registered and resampled respectively.

3. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 1 is characterized in that, In step S2, the spodumene spectrum index is calculated and the spodumene spectrum remote sensing index cube is prepared, including: S201, selecting the green band and the near infrared band of the pre-processed Landsat 8 image in the visible light-near infrared range to perform band ratio analysis to obtain the spodumene visible light-near infrared remote sensing index BR; S202, selecting the short-wave infrared band of the pre-processed GF-5 image, extracting the spectrum of the pixel where spodumene is located in the collected sample and averaging it as the reference spectrum of the short-wave infrared band of spodumene, and using the matched filter method to calculate the matched filter value MF between all pixels and the reference spectrum as the short-wave infrared remote sensing index of spodumene; S203, for the pre-processed SDGSAT-1 image, extract the three-band radiation brightness values ​​of the pixels where all spodumene samples are located, combine them two by two, and use the least squares method to fit the linear relationship of the thermal infrared radiation brightness of spodumene in each band combination; S204, combining the visible light-near infrared remote sensing index, the short-wave infrared remote sensing index and multiple thermal infrared remote sensing indexes corresponding to each pixel to obtain a three-dimensional spodumene remote sensing index cube, which is arranged in pixel order and stored in a two-dimensional csv format.

4. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 3 is characterized in that, In step S201, the calculation formula of spodumene visible light-near infrared remote sensing index BR is: In the formula, Band 3 is the green band and Band 5 is the near-infrared band.

5. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 3 is characterized in that, In step S202, the calculation formula of the matched filter value MF is: In the formula, S i is the actual spectral value of the pixel in the i-th band, R i is the reference spectrum value of spodumene in the i-th band, and n is the total number of bands.

6. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 3 is characterized in that, In step S202, the spodumene short-wave infrared remote sensing index ST i The calculation formula is: ST i =|Band b -Band b ×m i -c i | Where, Band a ,Band b are the two bands in the i-th band combination, m i ,c i They are the slope and intercept of the linear relationship obtained by linear fitting of spodumene in the th band combination.

7. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 1 is characterized in that, In step S3, a two-branch neural network model based on a graph convolutional network and a one-dimensional convolutional neural network is constructed, and the spatial distribution characteristics and spectral characteristics of spodumene are identified based on the two-branch neural network model, including: S301, use Matlab software to produce the data required for training; S302, build a deep learning model using Python language.

8. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 7 is characterized in that, In step S301, Matlab software is used to prepare the data required for training, including: (1) Read the image data and divide it into training set and test set according to the label of each pixel; (2) Perform data conversion to convert the three-dimensional remote sensing image data and label data into two-dimensional matrices respectively; (3) Calculate the Laplacian matrix required for graph convolutional networks; (4) Save the training and test data as mat files respectively.

9. The multi-source remote sensing spodumene identification method based on spectral index and double-branch neural network according to claim 7 is characterized in that, In step S302, a deep learning model is constructed using Python language, including: (a) Import the Python libraries of Numpy, Tensorflow, and Scipy; (b) Define auxiliary functions, including: sample_mask, creates a sample mask for data segmentation; create_placeholders, creates TensorFlow placeholders for input data and labels; initialize_parameters, initialize network parameters, including weights and biases; GCN_layer, defines the operation of the graph convolution layer, with the input being the feature matrix and the Laplacian matrix; Mynetwork, defines the forward propagation process of the hybrid model, which contains GCN layers and 1D-CNN; (c) Define the network structure, which consists of a graph convolution layer and a 1D-CNN. The graph convolution layer performs convolution operations on features through the Laplacian matrix, and the 1D-CNN is used to process the temporal information of the feature vector. Finally, the fully connected layer performs classification and outputs the probability distribution of the spodumene category. (d) Use cross entropy as the loss function, add L2 regularization, and use Adam optimizer for parameter update; (e) Use the leave-one-out cross-validation method to train the deep learning model, leaving one sample as the validation set each time and the rest as the training set. Then, train the deep learning model on the entire training set and make predictions on the test set to complete the spodumene identification. (f) Save the predicted probability of the test set as a MAT file and a CSV file for subsequent analysis.

Images