Deep mineral rapid detection method based on multiband spectrum fusion

Abstract

The invention relates to the technical field of deep mineral detection, and discloses a deep mineral rapid detection method based on multiband spectrum fusion, which comprises the following steps: step 1, building a detection system; 2, carrying the detection system on a detection platform, and synchronously collecting three types of spectral data and course angle, pitch angle, roll angle and speed parameters; step 3, data preprocessing; and step 4, a cross-modal fusion module realizes alignment of three types of data through cross-modal feature matching, and then weighted fusion is carried out. According to the deep mineral rapid detection method based on multi-band spectrum fusion, an attitude sensing module collects a course angle, a pitch angle, a roll angle and speed parameters in real time, a data preprocessing module cooperates with a cross-modal fusion module, and an attitude correction matrix is constructed based on Euler angle transformation to carry out geometric correction on multi-band spectrum data; the detection platform advances according to a preset height and speed working condition, a high-precision inertial navigation system is not needed, and the cost and the detection precision are both considered.

Description

Technical Field

[0001] This invention relates to the field of deep mineral exploration technology, specifically to a rapid deep mineral exploration method based on multi-band spectral fusion. Background Technology

[0002] Deep mineral exploration refers to exploration activities that use various detection technologies to obtain key information such as the distribution range, type, and reserves of mineral resources within a certain depth range underground. It is a core step before mineral resource development, providing data support for subsequent mining planning, cost control, and efficiency improvement, and is widely used in mineral resource exploration scenarios such as deep sea and deep land.

[0003] Current deep mineral exploration technologies are mainly divided into two categories: acoustic detection and optical detection. Acoustic detection: It has the characteristics of wide coverage and high imaging efficiency, and can realize the preliminary assessment of large-scale seabed topography and mineral distribution; Optical detection: It has high resolution and imaging accuracy, and can realize fine measurement of local areas. It mainly includes line structured light imaging technology and hyperspectral imaging technology. Line structured light technology projects a line beam and synthesizes a three-dimensional point cloud based on optical triangulation, which is suitable for fine detection at close range. Hyperspectral technology can acquire spatial information and spectral information at the same time, and distinguish mineral types by differences in spectral characteristics.

[0004] However, existing deep mineral exploration technologies still face the following pressing problems that need to be addressed: 1. Platform attitude and speed changes lead to imaging errors: Optical detection systems are usually mounted on platforms such as underwater vehicles and ground exploration vehicles. Unstable platform speed or attitude changes can cause large errors in the direction of travel in three-dimensional imaging. High-precision inertial navigation systems are expensive and have drift errors, making them difficult to widely apply. 2. Cross-modal data alignment is difficult: The three-dimensional point cloud generated by line structured light is an unordered set of coordinates, while hyperspectral data is an ordered three-dimensional matrix. The two data representations are significantly different, and the line scanning acquisition process is prone to introducing geometric deformation, which makes accurate alignment difficult and affects the information fusion effect. 3. Insufficient information abundance of hyperspectral data: In deep environments such as the deep sea, light propagation is strongly affected by absorption and scattering, resulting in severe distortion of hyperspectral data, especially in the red and violet bands, and a significant decrease in information abundance. Furthermore, the absorption characteristics of water bodies vary greatly in different regions, making it difficult to accurately correct and reducing recognition accuracy. Therefore, a rapid detection method for deep mineral resources based on multi-band spectral fusion was proposed. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a rapid deep mineral exploration method based on multi-band spectral fusion. By constructing a multi-module collaborative detection system, it integrates the acquisition of visible light, near-infrared, and short-wave infrared multi-band spectral data. Targeted data preprocessing, cross-modal alignment, and feature fusion classification schemes are designed to address the core technical problems of deep exploration. Ultimately, this achieves rapid and high-precision detection of deep minerals, solving the problems of imaging errors caused by platform attitude and velocity changes, difficulties in cross-modal data alignment, and insufficient information abundance of hyperspectral data in existing deep mineral exploration technologies.

[0006] (II) Technical Solution To achieve the aforementioned goal of rapid and high-precision detection of deep mineral deposits by building a multi-module collaborative detection system that integrates visible light, near-infrared, and short-wave infrared multi-band spectral data acquisition, and addressing the core technical challenges of deep mineral exploration, this invention designs targeted data preprocessing, cross-modal alignment, and feature fusion classification schemes. The invention provides the following technical solution: a rapid deep mineral exploration method based on multi-band spectral fusion, comprising: Step 1: Build a detection system. This system includes a multi-band spectral acquisition module, an attitude sensing module, a data preprocessing module, a cross-modal fusion module, and an identification output module. The multi-band spectral acquisition module includes a visible light spectral acquisition unit, a near-infrared spectral acquisition unit, and a short-wave infrared spectral acquisition unit. Step 2: Mount the detection system on the detection platform and control it to move in the target area according to the preset operating conditions, simultaneously collecting three types of spectral data and heading angle. Pitch angle Roll angle Speed ​​parameters ; Step 3: The data preprocessing module performs targeted noise reduction on the visible light spectral data, spectral smoothing on the near-infrared spectral data, and feature enhancement on the short-wave infrared spectral data. Step 4: The cross-modal fusion module aligns the three types of data through cross-modal feature matching, then constructs a feature fusion classification network to extract spectral and spatial features, and performs weighted fusion through an attention fusion mechanism; Step 5: The network outputs information on mineral type, distribution range, and confidence level. The identification output module then visualizes and stores the results.

[0007] Preferably, the sampling frequency of the multi-band spectral acquisition module is 150Hz, and the attitude sensing module adopts an inertial measurement unit with a sampling frequency of 200Hz. Detection platforms include underwater vehicles, ground-based exploration vehicles, or aerial drones, among which: The underwater vehicle has a preset altitude range of 1-5m and a preset speed range of 0.3-1 knots. The multi-band spectral acquisition unit and attitude sensing module are waterproof. The preset height range of the ground exploration vehicle is 0.5-3m, and the preset travel speed range is 0.3-1m / s; The preset altitude range for the aerial drone is 5-20m, and the preset flight speed range is 1-5m / s.

[0008] Preferably, the visible light spectral acquisition unit has a spectral range of 380-780nm, a spectral resolution of 2nm, and a spatial resolution of 3.36mm; The near-infrared spectral acquisition unit has a spectral range of 780-1000nm, a spectral resolution of 2nm, and a spatial resolution of 3.36mm. The shortwave infrared spectral acquisition unit has a spectral range of 1000-2500nm, a spectral resolution of 4nm, and a spatial resolution of 4.12mm. The optical axis angles of the three types of acquisition units are all between 30° and 45°, forming complementary wavelength coverage of key absorption bands of mineral characteristics, ensuring consistent spatial resolution.

[0009] Preferably, the data preprocessing process of the data preprocessing module is as follows: Step 1: Targeted noise reduction of visible light spectral data employs a median filtering algorithm with adaptive window adjustment. By dynamically switching the filter window size, it suppresses noise while preserving the subtle spectral characteristics of minerals. Step 2: The spectral smoothing algorithm for near-infrared spectral data calculates the average value through a sliding window, suppresses random noise in the near-infrared band, and preserves the characteristic absorption peaks of minerals. Step 3: Feature enhancement processing of shortwave infrared spectral data adopts an algorithm based on wavelet transform. By decomposing the spectral data, enhancing the high-frequency coefficients corresponding to mineral features, and suppressing background noise, mineral features are effectively separated from background interference.

[0010] Preferably, the cross-modal feature matching process of the cross-modal fusion module is as follows: By integrating principal component analysis, segmentation network, and feature matching network, cross-modal data alignment is achieved by extracting principal component features from three types of spectral data, obtaining mask images of mineral target areas, and optimizing the feature point matching cost matrix, thus meeting the accuracy requirements of subsequent fusion classification.

[0011] Preferably, the core design of the feature fusion classification network is as follows: The spectral feature branch employs a convolutional neural network, which enhances the ability to extract mineral spectral features through multi-layer convolution, activation functions, and normalization, thereby improving network convergence speed and avoiding gradient vanishing. The spatial feature branch uses a Transformer encoder, which contains multiple encoder layers. Each encoder layer integrates a multi-head self-attention mechanism and a feedforward neural network to improve the ability to capture the spatial structure features of mineral deposits. The formula for calculating the attention fusion mechanism is as follows: ,in For spectral characteristics, For spatial features, For attention weights, , and This is the weight matrix. For bias terms, It is the sigmoid activation function; The network training uses a multi-band spectral dataset containing various deep minerals. After deployment, the mineral identification accuracy is no less than 90%, and it supports batch processing of input data.

[0012] Preferably, the attitude correction process of the attitude sensing module is as follows: Step 1: Based on the Euler angle transformation formula An attitude correction matrix is ​​constructed to perform geometric correction on the acquired multi-band spectral data, thereby eliminating imaging deviations caused by changes in platform attitude. Step 2: The visualization display of the identification output module includes a mineral distribution heat map, a 3D imaging model, and a classification result statistics table. The storage formats are CSV and TIFF. The temperature gradient division interval of the heat map is no greater than 0.05, and the mesh accuracy of the 3D imaging model is no less than 1mm.

[0013] A rapid deep mineral exploration system based on multi-band spectral fusion is used for a rapid deep mineral exploration method based on multi-band spectral fusion. The system includes a multi-band spectral acquisition module, an attitude sensing module, a data preprocessing module, a cross-modal fusion module, and an identification output module. The multi-band spectral acquisition module includes a visible light spectral acquisition unit, a near-infrared spectral acquisition unit, and a short-wave infrared spectral acquisition unit. The three types of units are integrated into the same rigid bracket, with a fixed optical axis angle and connected in parallel through a waterproof data interface for synchronous acquisition of the three types of spectral data. The attitude sensing module is an inertial measurement unit that integrates a three-axis gyroscope and an accelerometer. It communicates bidirectionally with the data preprocessing module through a standardized data bus and outputs attitude parameters in real time. The data preprocessing module is a dedicated signal processing board that integrates noise reduction circuits, smoothing circuits, and feature enhancement circuits, and performs targeted hardware processing on the three types of spectral data respectively. The cross-modal fusion module is an embedded computing unit that embeds cross-modal feature matching logic and feature fusion classification network firmware to achieve data alignment and feature fusion. The identification output module includes a storage unit and a visualization interface, supporting the generation and storage of mineral distribution heat maps, three-dimensional imaging models, and classification result statistical tables; Each module is mechanically fixed by a waterproof connector and enables data interaction. They work collaboratively according to the process of "acquisition-preprocessing-alignment-fusion-output", have modular integration characteristics, and can be adapted to various detection platforms.

[0014] (III) Beneficial Effects Compared with existing technologies, this invention provides a rapid detection method for deep mineral resources based on multi-band spectral fusion, which has the following advantages: 1. This rapid deep mineral exploration method based on multi-band spectral fusion uses an attitude perception module to collect heading, pitch, roll, and velocity parameters in real time. A data preprocessing module, in conjunction with a cross-modal fusion module, constructs an attitude correction matrix based on Euler angle transformation to perform geometric correction on the multi-band spectral data. The exploration platform travels at a preset altitude and speed, eliminating the need for an expensive and high-precision inertial navigation system. This effectively eliminates imaging errors caused by changes in platform attitude and velocity, balancing cost and exploration accuracy.

[0015] 2. This rapid detection method for deep mineral resources based on multi-band spectral fusion integrates principal component analysis segmentation network and feature matching network in the cross-modal fusion module. First, it extracts principal component features of three types of spectral data to obtain the mask image of the mineral target area. Then, it optimizes the feature point matching cost matrix and combines it with standardized data after data preprocessing to solve the problems of cross-modal data expression form differences and geometric deformation, achieves accurate alignment, and significantly improves the information fusion effect.

[0016] 3. This rapid detection method for deep minerals based on multi-band spectral fusion uses the visible, near-infrared, and short-wave infrared units of the multi-band spectral acquisition module to complementarily cover the key absorption bands of mineral characteristics. The data preprocessing module improves data quality through targeted noise reduction, smoothing, and feature enhancement. The cross-modal fusion module uses a feature fusion classification network to extract spectral and spatial features, making up for the lack of information abundance of hyperspectral data in deep environments and significantly improving the accuracy of mineral identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of the rapid deep mineral exploration method of the present invention; Figure 2 This is a diagram of the architecture of the rapid deep mineral exploration system of the present invention; Figure 3 This is a detailed hardware diagram of the multi-band spectral acquisition module of the present invention; Figure 4 This is a flowchart of the cross-modal fusion module of the present invention; Figure 5 This is a schematic diagram of the attitude correction principle of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-5 A rapid detection method for deep mineral deposits based on multi-band spectral fusion, comprising: Step 1: Build a detection system. This system includes a multi-band spectral acquisition module, an attitude sensing module, a data preprocessing module, a cross-modal fusion module, and an identification output module. The multi-band spectral acquisition module includes a visible light spectral acquisition unit, a near-infrared spectral acquisition unit, and a short-wave infrared spectral acquisition unit. Step 2: Mount the detection system on the detection platform and control it to move in the target area according to the preset operating conditions, simultaneously collecting three types of spectral data and heading angle. Pitch angle Roll angle Speed ​​parameters ; Step 3: The data preprocessing module performs targeted noise reduction on the visible light spectral data, spectral smoothing on the near-infrared spectral data, and feature enhancement on the short-wave infrared spectral data. Step 4: The cross-modal fusion module aligns the three types of data through cross-modal feature matching, then constructs a feature fusion classification network to extract spectral and spatial features, and performs weighted fusion through an attention fusion mechanism; Step 5: The network outputs information on mineral type, distribution range, and confidence level. The identification output module then visualizes and stores the results.

[0020] A rapid deep mineral exploration system based on multi-band spectral fusion is used for a rapid deep mineral exploration method based on multi-band spectral fusion. The system includes a multi-band spectral acquisition module, an attitude sensing module, a data preprocessing module, a cross-modal fusion module, and an identification output module. The multi-band spectral acquisition module includes a visible light spectral acquisition unit, a near-infrared spectral acquisition unit, and a short-wave infrared spectral acquisition unit. The three types of units are integrated into the same rigid bracket, with a fixed optical axis angle and connected in parallel through a waterproof data interface for synchronous acquisition of the three types of spectral data. The attitude sensing module is an inertial measurement unit that integrates a three-axis gyroscope and an accelerometer. It communicates bidirectionally with the data preprocessing module through a standardized data bus and outputs attitude parameters in real time. The data preprocessing module is a dedicated signal processing board that integrates noise reduction circuits, smoothing circuits, and feature enhancement circuits, and performs targeted hardware processing on the three types of spectral data respectively. The cross-modal fusion module is an embedded computing unit that embeds cross-modal feature matching logic and feature fusion classification network firmware to achieve data alignment and feature fusion. The identification output module includes a storage unit and a visualization interface, supporting the generation and storage of mineral distribution heat maps, three-dimensional imaging models, and classification result statistical tables; Each module is mechanically fixed by a waterproof connector and enables data interaction. They work collaboratively according to the process of "acquisition-preprocessing-alignment-fusion-output", have modular integration characteristics, and can be adapted to various detection platforms.

[0021] Example 1: This embodiment focuses on the core process of rapid deep mineral exploration using multi-band spectral fusion, breaking down the entire operation from system setup to result output in detail to ensure the feasibility and high-precision output of the exploration method.

[0022] I. Detection System Setup: First, the selection and integration of each module were completed. In the multi-band spectral acquisition module, the visible light spectral acquisition unit selected a linear scan spectral sensor with a spectral range of 380-780nm, a spectral resolution of 2nm, and a spatial resolution of 3.36mm. The parameters of the near-infrared spectral acquisition unit were the same as those of the visible light unit. The short-wave infrared spectral acquisition unit selected a sensor with a spectral range of 1000-2500nm, a spectral resolution of 4nm, and a spatial resolution of 4.12mm. The three types of units were fixed on the same rigid bracket, and the optical axis angle was adjusted to 30° to ensure the consistency of spatial resolution. The attitude perception module uses an inertial measurement unit with a sampling frequency of 200Hz. This unit integrates a three-axis gyroscope and an accelerometer, and establishes bidirectional communication with the data preprocessing module through a standardized data bus. The data preprocessing module uses a dedicated signal processing board that integrates noise reduction circuits, smoothing circuits, and feature enhancement circuits to meet the processing needs of three types of spectral data. The cross-modal fusion module uses an embedded computing unit, which is written into the cross-modal feature matching logic and feature fusion classification network through firmware embedding. The identification output module is configured with a high-speed storage unit and a visualization interface. The storage unit interface type is SATA III, and the visualization interface supports HDMI output. It is pre-configured with CSV and TIFF format storage protocols.

[0023] II. Detection Platform Setup and Parameter Settings: A ground-based detection vehicle was selected as the detection platform. The platform was set to a preset height of 1m and a preset travel speed of 0.5m / s. The completed detection system was fixed to the top of the platform using mechanical connectors to ensure the stability of the system's center of gravity and avoid additional vibrations during travel.

[0024] III. Data Collection: The detection system is activated, controlling the ground detection vehicle to move within the target area according to preset operating conditions, simultaneously triggering the multi-band spectral acquisition module and attitude sensing module; the three types of spectral acquisition units simultaneously acquire visible light, near-infrared, and short-wave infrared spectral data of the target area at a sampling frequency of 150Hz; the attitude sensing module acquires the heading angle in real time. Pitch angle Roll angle Speed ​​parameters The collected data is transmitted to the data preprocessing module via a waterproof connector, with the transmission delay controlled within 10ms. IV. Data Preprocessing: The data preprocessing module processes the three types of spectral data according to a preset process; Visible light spectral data: An adaptive window-adjusted median filtering algorithm is used to calculate the noise variance of local spectral data through a sliding window. The initial window size is set to 3×3; noise variance The calculation follows the formula ,in For a single spectral data point within the window, Let n be the mean of the data within the window, and n be the number of data points within the window; when When the value is greater than 0.05, the window dynamically switches to a 5x5 window. When the value is ≤0.05, a 3×3 window is maintained to preserve the subtle spectral features of the mineral while suppressing random noise; Near-infrared spectral data: A sliding window smoothing algorithm is used, with the window size set to 7 consecutive sampling points. The average value of the spectral data within the window is calculated point by point to suppress random noise in the near-infrared band and ensure the integrity of the mineral characteristic absorption peaks. Shortwave infrared spectral data: A wavelet transform algorithm based on the db4 wavelet basis was used to decompose the spectral data into three layers. The high-frequency coefficients corresponding to mineral features (the detail coefficients of the third layer) were multiplied by an enhancement factor of 1.5, and the low-frequency coefficients corresponding to background noise were multiplied by an attenuation factor of 0.8. The spectral data was then reconstructed through inverse wavelet transform to effectively separate mineral features from background interference. The wavelet transform decomposition formula is as follows: ,in This is the raw shortwave infrared spectral data. The coefficients are for three levels of approximation. , , These are the detail coefficients for three layers: horizontal, vertical, and diagonal; the reconstruction formula is... ,in For the reconstructed spectral data, This represents the inverse wavelet transform operator.

[0025] V. Cross-modal fusion: The cross-modal fusion module completes data processing according to the following steps, the core of which includes the construction, training and deployment of the feature matching network and the fusion classification network.

[0026] Cross-modal feature matching: Network Construction: The principal component analysis algorithm is integrated with the U-Net segmentation network and feature matching network to form a cross-modal feature matching architecture. The principal component analysis algorithm is used for data dimensionality reduction. The U-Net segmentation network consists of 3 layers each for the encoder and decoder. The encoder adopts a structure of convolutional layers plus max pooling layers, and the decoder adopts a structure of deconvolutional layers plus sampling layers. The feature matching network consists of a feature extraction subnetwork and a cost matrix optimization subnetwork. Training process: The U-Net segmentation network was trained using a multi-band spectral dataset containing 10 types of deep minerals. The dataset contained 5000 labeled images, which were divided into training and validation sets in an 8:2 ratio. The training epochs were set to 30, the optimizer was Adam, the learning rate was set to 1e-4, and the loss function was cross-entropy loss. The feature matching network generated training samples by simulating the distribution of feature points in cross-modal data. The number of training samples was 8000. During the training process, the network parameters were optimized with the goal of minimizing the feature point matching error. Deployment and Operation: The trained model is embedded into the computing unit. During runtime, principal component analysis (PCA) is first used to reduce the dimensionality of the three types of spectral data, extracting the top three principal components with a cumulative contribution rate of 95%. The core formula for PCA is Y=PX, where X is the original spectral data matrix, P is the eigenvector matrix, and Y is the dimensionality-reduced feature matrix. Then, a U-Net segmentation network is used to segment the dimensionality-reduced feature map, obtaining a binary mask image of the mineral target area to eliminate data morphology differences. Finally, a feature matching network is used to calculate the Euclidean distance between feature points. The Euclidean distance formula is... ,in , Let be the k-th eigenvalues ​​of the two modalities, and m be the feature dimension. An initial cost matrix is ​​constructed based on Euclidean distance, and the matrix is ​​optimized using the Hungarian algorithm to complete cross-modal data alignment, with the alignment error controlled within 0.1 pixels.

[0027] Feature fusion and classification: Network Construction: The feature fusion classification network includes a spectral feature branch, a spatial feature branch, and an attention fusion mechanism. The spectral feature branch uses three convolutional layers with kernel sizes of 3×3, numbered 16, 32, and 64 respectively, with a stride of 1 and padding of 1. Each convolutional layer is followed by a BatchNorm normalization layer, using ReLU as the activation function. The spatial feature branch uses two encoder layers, each integrating an 8-head self-attention mechanism and a feedforward neural network. The feedforward neural network has a hidden layer dimension of 256. The core formula of the attention fusion mechanism is: Where F represents the final fusion feature. The output features are for the spectral feature branch. Output features for the spatial feature branch. For attention weights, and This is the weight matrix. For bias terms, The activation function is sigmoid; the weight matrix is... and Xavier uniform distribution initialization is used, with the bias term b set to 0.1. The value range is 0.3-0.7; Training Process: A multi-band spectral dataset containing 10,000 samples from 10 deep mineral deposits was used for training. The dataset was divided into training and validation sets in an 8:2 ratio, with a batch size of 32. The training run consisted of 50 epochs, and cross-entropy loss was used as the loss function. The cross-entropy loss formula is as follows: ,in For the number of mineral categories, For the true labels of the samples, To predict probabilities for the model, the optimizer was Adam, and the learning rate was set to 1e-4. During training, an early stopping strategy was used to prevent overfitting, and training was stopped when the validation set loss did not decrease for 5 consecutive rounds. Deployment and operation: The trained network model is embedded in the embedded computing unit of the cross-modal fusion module, which supports batch processing of input data. During the processing, the spectral feature branch and spatial feature branch are called in real time to extract corresponding features. The weights are dynamically calculated and feature fusion is completed through the attention fusion mechanism.

[0028] VI. Attitude Correction and Result Output: The attitude perception module constructs the attitude correction matrix based on the Euler angle transformation formula, which is: in This is the attitude correction matrix. Let be the rotation matrix about the z-axis. Let be the rotation matrix about the y-axis. Let be the rotation matrix about the x-axis. For heading angle, The pitch angle, This refers to the roll angle; The matrix is ​​used to perform geometric correction on the aligned spectral data to eliminate imaging deviations caused by attitude changes during platform movement. The recognition output module generates a mineral distribution heat map, a 3D imaging model, and a classification result statistics table. The temperature gradient division interval of the heat map is 0.05, and the mesh accuracy of the 3D imaging model is 1mm. The data is stored as classification results in CSV format and heat map and 3D model in TIFF format through the SATA III interface, and visualized in real time through the HDMI interface.

[0029] In this embodiment, the spectral ranges of the three types of spectral acquisition units complement each other, covering the key absorption bands of mineral features. The optical axis angle design ensures consistent spatial resolution, providing a data foundation for accurate identification. Targeted preprocessing algorithms address visible light noise, near-infrared random interference, and short-wave infrared background interference, improving data quality. Cross-modal feature matching achieves accurate alignment through principal component analysis dimensionality reduction, segmentation network masking, and cost matrix optimization, laying the foundation for feature fusion. The feature fusion classification network simultaneously extracts spectral and spatial features. Spectral features reflect the chemical composition of the mineral, while spatial features reflect the morphological distribution of the mineral. An attention mechanism dynamically balances the contribution weights of the two types of features, and attitude correction eliminates imaging bias, ultimately achieving rapid and efficient detection.

[0030] Example 2: This embodiment addresses the environmental differences of three types of detection platforms: underwater vehicles, ground-based exploration vehicles, and aerial drones. It designs specific adaptation solutions, clarifies the parameter adjustment, hardware modification, and data processing optimization logic for each platform, and ensures the stable operation and accurate output of the detection system in different scenarios.

[0031] I. Underwater Vehicle Compatibility: Platform parameter settings: preset altitude 3m, preset speed 0.6 knots (approximately 0.309m / s), the multi-band spectrum acquisition module and attitude perception module are waterproofed with an IP68 waterproof rating, and the data interface uses an M12 waterproof connector to avoid seawater corrosion. Data acquisition adaptation: The inertial measurement unit monitors the platform attitude changes caused by water flow in real time and outputs attitude parameters every 5ms. The acquisition deviation of the spectral data is corrected by the compensation algorithm of the data preprocessing module. The trigger frequency of the multi-band spectral acquisition unit is synchronized with the sampling frequency of the attitude sensing module to achieve frequency division synchronization of 150Hz and 200Hz to ensure that the timestamps of the spectral data and attitude parameters are aligned. Special processing: Due to the severe attenuation of light in the underwater environment, the integration time of the shortwave infrared spectroscopy acquisition unit was adjusted to 50μs to improve signal strength. At the same time, the feature enhancement circuit of the data preprocessing module was used to further enhance the mineral characteristic signal.

[0032] II. Ground exploration vehicle compatibility: Platform parameter settings: preset height 2m, preset driving speed 0.8m / s. For complex terrain with a slope of ≤15°, the acceleration measurement range of the attitude perception module is set to ±2g to improve the sensitivity to ground bumps. Data acquisition and adaptation: The platform's shock absorption device reduces vibration interference during driving, and the noise reduction circuit of the data preprocessing module filters low-frequency noise below 10Hz to ensure the stability of spectral data. Motion control: The GPS positioning-assisted platform moves along a preset path with a positioning accuracy of ≤1m to ensure full coverage of the detection area.

[0033] III. Aerial Drone Compatibility: Platform parameter settings: preset altitude 10m, preset flight speed 3m / s, and an anti-fog coating is added to the lens of the multi-band spectral acquisition unit to avoid image blurring caused by high-altitude water vapor; Data acquisition adaptation: The flight speed of the UAV is linked to the trigger frequency of the spectral acquisition unit through the flight control system. When the flight speed increases to 5m / s, the sampling frequency of the spectral acquisition unit remains unchanged at 150Hz. The spatial resolution loss caused by the increase in speed is compensated by increasing the number of pixels acquired in a single acquisition. The attitude perception module combines the gyroscope data of the UAV to correct the spectral imaging deviation caused by changes in flight attitude in real time. Data transmission: A 5G wireless transmission module is used to transmit the collected data to the ground receiving end in real time, with a transmission delay of ≤50ms to ensure the real-time performance of data processing.

[0034] In this embodiment, specific adaptations are made for the environmental characteristics of different detection platforms. Underwater platforms focus on waterproof encapsulation and light compensation, ground platforms are enhanced with shock absorption and low-frequency noise filtering, and aerial platforms focus on anti-fog treatment and speed adaptation. Through parameter adjustment, hardware modification, and algorithm compensation, the detection system can be made to work stably on various platforms. The preset altitude and speed range of the three types of platforms meet the adaptation requirements of the detection method. Through synchronous control and attitude correction, the accuracy of spectral data and attitude parameters is guaranteed, which significantly improves the adaptability of the detection system and effectively expands the application scope of the technical solution.

[0035] Example 3: This embodiment focuses on the preprocessing requirements of three types of spectral data, and breaks down in detail the construction logic, training process, deployment parameters and operation flow of three algorithms: adaptive window mid-value filtering, weighted sliding window smoothing and wavelet transform enhancement, to ensure that the quality of the preprocessed data meets the requirements of subsequent fusion and classification.

[0036] I. Targeted noise reduction processing for visible light spectral data (median filtering algorithm with adaptive window adjustment): Algorithm Construction: With the core objective of dynamically suppressing noise and preserving subtle features, an adaptive switching logic for window size is designed. The window size includes three categories: 3×3, 4×4, and 5×5. The switching is based on the local noise variance threshold.

[0037] Training and Calibration: The algorithm was calibrated using a visible light spectrum dataset containing different noise intensities. The dataset covers 1000 sets of spectral data from mineral-bearing scenes. By labeling the noise variance and feature retention rate, the variance thresholds σ1=0.03 and σ2=0.05 were determined. When the noise variance σ>σ2, the window was switched to 5×5; when σ1<σ≤σ2, the window was switched to 4×4; and when σ≤σ1, the window was kept at 3×3.

[0038] Deployment and Operation: Step 1: Normalize the input visible light spectrum data to map the data range to the [0,1] interval to reduce processing bias caused by differences in data magnitude; Step 2: Construct a 3×3 sliding window, traverse the entire visible light spectral data matrix, and apply the formula... Calculate the noise variance of the data within each window. ,in For a single data point within the window, This represents the average of the data within the window. The number of data items in the window; Step 3: Based on noise variance The sliding window size is dynamically adjusted based on the comparison results with a preset threshold. Step 4: Sort the data in each window by median, and take the median value as the output value of the center pixel of the window to complete the filtering process; Step 5: Verify the processing effect using the peak signal-to-noise ratio (PSNR). The PNR formula is: MAX is the maximum gray value of the data, MSE is the mean square error of the data before and after processing, and PSNR is required to be ≥35dB to ensure noise suppression while preserving the subtle spectral characteristics of the mineral.

[0039] II. Spectral smoothing of near-infrared spectral data (weighted sliding window smoothing algorithm): Algorithm Construction: A sliding window with 9 consecutive sampling points is designed, with a window step size of 1 sampling point. The weight of each sampling point in the window is allocated using a weighted average logic, with the center sampling point having the highest weight to reduce interference from edge data.

[0040] Weight calibration: Through testing with multiple sets of near-infrared spectral data, the weight of the sampling point at the center of the window was determined to be 0.3, the weight of each of the two adjacent sampling points was 0.15, and the weight of each of the remaining sampling points was 0.1, to ensure that the shape of the characteristic absorption peaks is not distorted after smoothing.

[0041] Deployment and Operation: Step 1: Set the sliding window size to 9 consecutive sampling points and the window step size to 1 sampling point; Step 2: Perform a weighted average calculation on the spectral data within the window according to preset weights. The weighted average formula is as follows: ,in Output the value to the center of the window. For the first The weight of each sampling point For the first Spectral data from each sampling point; Step 3: Traverse the entire near-infrared spectral data, calculate the window-weighted average point by point, and replace the corresponding points in the original data to achieve spectral smoothing; Step 4: Verify the processed data, requiring the absolute value of the first derivative of the spectral curve to be ≤0.02 (except for the mineral characteristic absorption peak region), to ensure that random noise is suppressed while the shape of the mineral characteristic absorption peak is preserved.

[0042] III. Feature Enhancement Processing of Shortwave Infrared Spectroscopic Data (Based on Wavelet Transform Algorithm): Algorithm Construction: The db4 wavelet basis is selected as the decomposition basis function, and a 3-level decomposition hierarchy is set. The high-frequency coefficients corresponding to mineral characteristics are strengthened and the low-frequency coefficients corresponding to background noise are suppressed through the coefficient adjustment strategy.

[0043] Coefficient calibration: By analyzing the short-wave infrared spectral characteristics of different types of minerals, the enhancement factor of the high-frequency coefficient k=1.5 and the attenuation factor of the low-frequency coefficient are determined to be 0.8, so as to ensure effective separation of features and noise.

[0044] Deployment and Operation: Step 1: Select the db4 wavelet basis to perform a 3-level wavelet decomposition on the shortwave infrared spectral data to obtain one approximation coefficient. and 3 detail coefficients , , The decomposition formula is ,in This is the raw spectral data; Step 2: Identify the frequency range corresponding to the mineral characteristics and determine... , , The high-frequency coefficients in the figure are the correlation coefficients of mineral characteristics. These high-frequency coefficients are multiplied by an enhancement factor of 1.5, and the low-frequency coefficients corresponding to the background noise are... The components in the formula are multiplied by an attenuation factor of 0.8; Step 3: Using the inverse wavelet transform algorithm, the processed approximation coefficients and detail coefficients are reconstructed to obtain the feature-enhanced shortwave infrared spectral data. The reconstruction formula is as follows: ; in For the reconstructed data, It is the inverse wavelet transform operator; Step 4: Verify the effect by calculating the signal-to-noise ratio (SNR) of the data before and after processing. The SNR formula is: ,in For signal power, For noise power, the requirements are: Improve by ≥10dB to ensure effective separation of mineral characteristics from background interference.

[0045] In this embodiment, the adaptive windowed value filtering of visible light spectral data dynamically adjusts the window size according to the noise intensity, avoiding the problems of insufficient noise suppression or feature loss caused by a fixed window, and better preserving subtle mineral spectral features; the weighted sliding window smoothing algorithm of near-infrared spectroscopy suppresses random noise while reducing the blurring of feature absorption peaks and ensuring the integrity of absorption peak shapes; the wavelet transform enhancement algorithm of short-wave infrared spectroscopy strengthens mineral feature signals in a targeted manner through decomposition and coefficient adjustment, improving the recognizability of mineral features; the three types of preprocessing algorithms work together to improve the effective information utilization rate of the original spectral data, providing high-quality data support for subsequent cross-modal fusion and classification.

[0046] Example 4: This embodiment breaks down the complete construction process of a cross-modal feature matching and feature fusion classification network in detail, from network architecture design and training process optimization to deployment and operation parameters, clarifying the technical details of each step to ensure accurate alignment and efficient fusion classification of cross-modal data.

[0047] I. Cross-modal feature matching: Network Construction: Principal Component Analysis (PCA), SegNet segmentation network, and Siamese feature matching network are integrated to form a three-level architecture of "dimensionality reduction-segmentation-matching". PCA is used for data dimensionality compression. SegNet segmentation network consists of 3 layers each for encoder and decoder. The encoder adopts a 3-layer convolution + max pooling structure, and the decoder adopts a 3-layer deconvolution + upsampling structure. Siamese feature matching network consists of two symmetrical feature extraction branches and a matching cost calculation layer.

[0048] Training process: Dataset preparation: Construct a training set containing 10,000 sets of cross-modal spectral data, covering different mineral types and environmental backgrounds, and divide it into training set and validation set in an 8:2 ratio; Principal Component Analysis Calibration: Principal component analysis was performed on the three types of spectral data in the training set. The top three principal components with a cumulative contribution rate of 95% were identified as the dimensionality reduction targets. The dimensionality reduction formula is Y=PX, where X is the original data matrix, P is the eigenvector matrix, and Y is the dimensionality-reduced eigenvector matrix. SegNet training: The goal is to optimize the segmentation accuracy of mineral target regions. The cross-entropy loss function is used, and the loss formula is as follows: ,in This is the mask label (1 for the target area, 0 for the background area). To predict probabilities, the optimizer was Adam, the learning rate was set to 1e-4, and the training epochs were set to 50 epochs to ensure that the segmented binary mask image could accurately delineate the target mineral area. Siamese network training: The optimization objective is feature point matching accuracy, using a contrastive loss function. The loss formula is as follows: ,in For the sample size, To match the tag (1 for match, 0 for no match). For feature point descriptors of two modalities, The marginal parameter is set to 2.0, the number of training rounds is set to 60, and the learning rate is dynamically adjusted from 1e-4 to 1e-5.

[0049] Deployment and Operation: Step 1: Principal component analysis is performed to reduce the dimensionality of the three types of spectral data after preprocessing and extract the top three principal components with a cumulative contribution rate of 95% to reduce data redundancy. Step 2: Segmentation network processing. Input the dimensionality-reduced feature matrix into the SegNet segmentation network and output a binary mask image, in which the mineral target area is marked as 1 and the background area is marked as 0, to eliminate data morphology differences. Step 3: Feature matching network processing. The masked images of the three types of spectral data are input into the Siamese network to extract feature point descriptors. The Euclidean distance between feature points of different modes is calculated. The Euclidean distance formula is... ,in For feature dimension, An initial cost matrix is ​​constructed for the feature point coordinates of different modalities. The matrix is ​​optimized using the Hungarian algorithm to minimize the matching cost and complete cross-modal data alignment. The feature point matching accuracy after alignment is ≥95%, and the alignment error is controlled within 0.1 pixels.

[0050] II. Feature Fusion Classification Network: Network Construction: This includes a spectral feature branch, a spatial feature branch, and an attention fusion mechanism. The spectral feature branch employs a convolutional neural network architecture, consisting of 3 convolutional layers, 3 activation layers, 3 normalization layers, and 1 pooling layer. The first convolutional layer has a kernel size of 3×3 and 16 kernels; the second has a kernel size of 3×3 and 32 kernels; and the third has a kernel size of 3×3 and 64 kernels. All layers have a stride of 1 and padding of 1. The activation function is ReLU, the normalization layer uses BatchNorm, and the pooling layer uses max pooling (2×2 window, stride 2). The spatial feature branch uses a Transformer encoder architecture, containing two encoder layers. Each encoder layer integrates an 8-head self-attention mechanism and a feedforward neural network. The self-attention mechanism has a query Q, key K, and value V dimension of 64. The feedforward neural network contains two fully connected layers with a hidden layer dimension of 256 and an output dimension of 64. The core formula for the attention fusion mechanism is: in As a feature of fusion, This is the 64-dimensional feature vector output by the spectral feature branch. This is the 64-dimensional feature vector output by the spatial feature branch. For attention weights, and All are 64×64 weight matrices, and b is the bias term (set to 0.1). The sigmoid activation function is given by: .

[0051] Training process: Dataset preparation: A multi-band spectral dataset containing 10 kinds of deep minerals is used. The dataset contains 10,000 samples, which are divided into training set and validation set in an 8:2 ratio. The samples cover different mineral forms, distribution density and environmental background. Network training: Batch size is set to 32, training epochs are set to 50, and cross-entropy loss is used as the loss function. The formula is as follows: ,in For the number of mineral categories, For the true labels of the samples, To predict probabilities for the model, the optimizer was Adam, and the learning rate was set to 1e-4. During training, a learning rate decay strategy was adopted, with the learning rate being halved every 10 rounds. Model optimization: Overfitting is suppressed by using a dropout layer with a dropout probability of 0.2, added before the fully connected layer. An early stopping strategy is also adopted, stopping training when the validation set accuracy does not improve for 5 consecutive rounds. Deployment and Operation: The trained network model is deployed to the embedded computing unit and stable operation is achieved through firmware fixation. During operation, the spectral feature branch extracts features related to the chemical composition of minerals, the spatial feature branch captures features related to the morphological distribution of minerals, and the attention fusion mechanism dynamically calculates the α value to balance the contribution weights of the two types of features, and finally outputs the mineral classification results.

[0052] In this embodiment, cross-modal feature matching reduces data redundancy through principal component analysis for dimensionality reduction, segmentation network obtains accurate mask images to eliminate morphological differences, and feature matching network optimizes the cost matrix to achieve high-precision alignment, providing a precise data foundation for feature fusion. The convolutional neural network architecture of the spectral feature branch efficiently extracts the spectral features of minerals, while the Transformer encoder of the spatial feature branch captures the spatial structural features of minerals. The attention fusion mechanism dynamically balances the contributions of the two types of features, enabling the network to fully utilize the multidimensional information of minerals and achieve the dual requirements of high-precision identification and rapid detection.

[0053] In summary, this rapid deep mineral exploration method based on multi-band spectral fusion utilizes an attitude sensing module to collect real-time heading, pitch, roll, and velocity parameters. A data preprocessing module, in conjunction with a cross-modal fusion module, constructs an attitude correction matrix based on Euler angle transformation to perform geometric correction on the multi-band spectral data. The exploration platform travels at a preset altitude and speed, eliminating the need for an expensive, high-precision inertial navigation system. This effectively eliminates imaging errors caused by changes in platform attitude and velocity, balancing cost and exploration accuracy.

[0054] Furthermore, this rapid deep mineral exploration method based on multi-band spectral fusion integrates principal component analysis segmentation network and feature matching network in its cross-modal fusion module. It first extracts principal component features from three types of spectral data to obtain a mask image of the mineral target area, then optimizes the feature point matching cost matrix, and combines it with standardized data after data preprocessing to solve the problems of differences in cross-modal data expression forms and geometric deformation, achieving accurate alignment and significantly improving the information fusion effect.

[0055] Furthermore, this rapid deep mineral exploration method based on multi-band spectral fusion utilizes the complementary coverage of key absorption bands of mineral characteristics by the visible, near-infrared, and short-wave infrared units of the multi-band spectral acquisition module. The data preprocessing module improves data quality through targeted noise reduction, smoothing, and feature enhancement. The cross-modal fusion module extracts spectral and spatial features through a feature fusion classification network, compensating for the lack of information abundance in hyperspectral data in deep environments and significantly improving mineral identification accuracy. This method solves the problems of imaging errors caused by platform attitude and velocity changes, difficulties in cross-modal data alignment, and insufficient information abundance in hyperspectral data in existing deep mineral exploration technologies.

[0056] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid detection of deep mineral resources based on multi-band spectral fusion, characterized in that, The application relates to a mineral resource detection method based on multi-spectrum fusion. Step 1: a detection system is built, the system comprises a multi-band spectrum acquisition module, an attitude sensing module, a data preprocessing module, a cross-modal fusion module and an identification output module, the multi-band spectrum acquisition module comprises a visible light spectrum acquisition unit, a near-infrared spectrum acquisition unit and a short-wave infrared spectrum acquisition unit; Step 2: The detection system is carried on the detection platform, and the detection system is controlled to travel in the target area according to the preset working condition, and three types of spectral data and a heading angle are synchronously collected , a pitch angle , a roll angle , and a speed parameter ; Step 3: the data preprocessing module performs targeted noise reduction processing on the visible light spectrum data, performs spectrum smoothing processing on the near-infrared spectrum data and performs feature enhancement processing on the short-wave infrared spectrum data; Step 4: the cross-modal fusion module realizes alignment of the three types of data through cross-modal feature matching, then constructs a feature fusion classification network to extract spectrum features and space features, and realizes weighted fusion through an attention fusion mechanism; Step 5: the network outputs mineral resource types, distribution ranges and confidence information, and the identification output module visualizes and stores the results.

2. The method according to claim 1, characterized in that, The sampling frequency of the multi-band spectrum acquisition module is 150Hz, and the attitude sensing module adopts an inertial measurement unit and has a sampling frequency of 200Hz; The detection platform comprises an underwater vehicle, a ground detection vehicle or an aerial unmanned aerial vehicle, wherein: The preset height range of the underwater vehicle is 1-5m, the preset speed range is 0.3-1 knot, and the multi-band spectrum acquisition unit and the attitude sensing module have waterproof protection capability; The preset height range of the ground detection vehicle is 0.5-3m, and the preset driving speed range is 0.3-1m / s; The preset height range of the aerial unmanned aerial vehicle is 5-20m, and the preset flight speed range is 1-5m / s.

3. The method according to claim 1, characterized in that, The spectrum range of the visible light spectrum acquisition unit is 380-780nm, the spectrum resolution is 2nm, and the space resolution is 3.36mm; The spectrum range of the near-infrared spectrum acquisition unit is 780-1000nm, the spectrum resolution is 2nm, and the space resolution is 3.36mm; The spectrum range of the short-wave infrared spectrum acquisition unit is 1000-2500nm, the spectrum resolution is 4nm, and the space resolution is 4.12mm; The optical axis included angle of the three types of acquisition units is 30-45 degrees, the wavelength is complementary, the key bands of mineral resource characteristic absorption are covered, and the space resolution consistency is guaranteed.

4. The method according to claim 1, characterized in that, The data preprocessing process of the data preprocessing module is as follows: Step one, the targeted noise reduction processing of the visible light spectrum data adopts a median filtering algorithm with adaptive window adjustment, the filtering window size is dynamically switched, the noise is suppressed, and the fine spectrum features of the mineral resource are reserved; Step two, the spectrum smoothing algorithm of the near-infrared spectrum data calculates the average value through a sliding window, random noise in the near-infrared band is suppressed, and the mineral resource characteristic absorption peak is reserved; Step three, the feature enhancement processing of the short-wave infrared spectrum data adopts an algorithm based on wavelet transform, the spectrum data is decomposed, the high-frequency coefficient corresponding to the mineral resource feature is strengthened, and the background noise is suppressed, so that the mineral resource feature and the background interference are effectively separated.

5. The method according to claim 1, characterized in that, The cross-modal feature matching process of the cross-modal fusion module is as follows: The principal component analysis, the segmentation network and the feature matching network are integrated, the principal component features of the three types of spectrum data are extracted, the mineral resource target area mask image is obtained, the feature point matching cost matrix is optimized, the cross-modal data alignment is completed, and the subsequent fusion classification precision requirement is met.

6. The method according to claim 1, characterized in that, The core design of the feature fusion classification network is as follows: The spectral feature branch adopts a convolutional neural network, which strengthens the ability of mineral spectral feature extraction, improves the network convergence speed, and avoids gradient disappearance through multi-layer convolution, activation function and normalization processing; The spatial feature branch adopts a Transformer encoder, which contains multiple encoder layers, each of which integrates multi-head self-attention mechanism and feedforward neural network to improve the ability to capture the spatial structure features of minerals; The attention fusion mechanism calculation formula is wherein is a spectral feature, is a spatial feature, is an attention weight, , and is a weight matrix, is a bias term, is a sigmoid activation function; The network training uses a multi-band spectral dataset containing various deep minerals, and the mineral recognition accuracy after deployment is not less than 90%, supporting batch processing of input data.

7. The method according to claim 1, characterized in that, The posture correction process of the posture perception module is as follows: Step one, based on Euler angle transformation formula The pose correction matrix is constructed, multi-band spectral data collected is geometrically corrected, and imaging deviation caused by platform pose change is eliminated. Step two, the visual display of the identification output module includes a mineral distribution heat map, a three-dimensional imaging model and a classification result statistical table, and the storage format is CSV and TIFF, the temperature gradient division interval of the heat map is not greater than 0.05, and the grid accuracy of the three-dimensional imaging model is not less than 1mm.

8. A deep mineral rapid detection system based on multi-band spectral fusion, characterized in that, The method for realizing the multi-band spectral fusion-based rapid detection of deep minerals according to claims 1-7 comprises a multi-band spectral acquisition module, a posture perception module, a data preprocessing module, a cross-modal fusion module and an identification output module; The multi-band spectral acquisition module includes a visible light spectrum acquisition unit, a near-infrared spectrum acquisition unit and a short-wave infrared spectrum acquisition unit, which are integrated in the same rigid support, the optical axis angle is fixed and connected in parallel through a waterproof data interface, and used for synchronous acquisition of three types of spectral data; The posture perception module is an inertial measurement unit integrating a three-axis gyroscope and an accelerometer, which communicates with the data preprocessing module through a standardized data bus and outputs posture parameters in real time; The data preprocessing module is a special signal processing board, which integrates noise reduction circuit, smoothing processing circuit and feature enhancement circuit, and performs targeted hardware processing on three types of spectral data respectively; The cross-modal fusion module is an embedded computing unit, which solidifies the cross-modal feature matching logic and the feature fusion classification network firmware, realizes data alignment and feature fusion; The identification output module includes a storage unit and a visualization interface, which supports the generation and storage of mineral distribution heat map, three-dimensional imaging model and classification result statistical table; Each module is mechanically fixed through a waterproof connector and realizes data interaction, and works collaboratively according to the "acquisition-preprocessing-alignment-fusion-output" process, has the characteristics of modular integration, and can be adapted to various detection platforms.

Images