Rock weathering intelligent detection method based on multispectral imaging technology

Abstract

The invention discloses an intelligent rock weathering detection method based on a multispectral imaging technology, and particularly relates to the field of rock detection.The method comprises the steps that a multispectral imaging device is used for collecting a rock sample image, and then radiometric calibration, geometric correction and image enhancement processing are sequentially conducted on the image. The method comprises the following steps of: preprocessing an image, acquiring spectrum, texture and morphological characteristic parameters from the preprocessed image, collecting samples through a multi-source database, constructing a deep convolutional neural network model, inputting the acquired characteristic parameters into the model to obtain a rock weathering probability vector, judging the weathering degree according to a maximum probability criterion, and outputting a result to a user side. According to the method, multispectral imaging and deep learning technologies are fused, the defects of a traditional detection method are overcome, the rock weathering degree can be detected more accurately and efficiently, nondestructive detection and real-time monitoring are achieved, and the method has important application value in the fields of geological research, engineering construction and the like.

Description

Technical Field

[0001] The present invention relates to the field of rock detection technology, and more specifically, to an intelligent rock weathering detection method based on multi-spectral imaging technology. Background Art

[0002] In the field of rock weathering detection, accurately judging the degree of rock weathering is of great significance to many fields. Currently, there are mainly traditional detection methods and existing detection methods based on multispectral imaging technology.

[0003] Traditional rock weathering detection methods include on-site observation, chemical analysis, and physical testing. The on-site observation method mainly relies on the naked eye to observe the appearance characteristics of rocks, such as color changes, surface texture, crack development, etc., to infer the degree of rock weathering. The chemical analysis method is to collect rock samples and analyze their chemical composition in the laboratory to determine the degree of weathering by changes in composition. The physical testing method is to use ultrasonic testing, density testing and other means to evaluate the degree of weathering based on changes in the physical properties of rocks. With the development of technology, multispectral imaging technology has been applied in rock weathering detection. It uses the reflection, absorption and scattering characteristics of light in different bands after interacting with rocks to obtain the spectral information and spatial information of rocks in multiple narrow bands, providing new perspectives and data support for rock weathering detection from a spectral perspective.

[0004] However, there are still some shortcomings in its actual use. For example, the existing detection methods based on multispectral imaging technology are not comprehensive enough in feature extraction. Most of them only focus on spectral features and fail to fully explore other important information such as texture and morphology contained in multispectral images, resulting in incomplete description of rock weathering characteristics. The intelligent detection model constructed by it is insufficient in accuracy and reliability. Since the training samples may not be rich and diverse enough, the generalization ability of the model is weak. It is difficult to accurately judge the weathering degree of rock samples under complex and changeable geological conditions. In addition, most of these methods lack the function of real-time monitoring, cannot track the dynamic changes of rock weathering degree in time, and cannot meet the demand for real-time information on rock weathering in practical applications. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a rock weathering intelligent detection method based on multispectral imaging technology, which solves the problems raised in the above-mentioned background technology through the following scheme.

[0006] To achieve the above-mentioned object, the present invention provides the following technical solutions: A method for intelligent detection of rock weathering based on multispectral imaging technology, comprising: S1: multispectral image acquisition: obtaining a rock sample image through the acquisition terminal by inputting preset acquisition parameters into the acquisition terminal; S2: Multispectral image preprocessing: The collected rock sample images are sequentially subjected to radiation calibration, geometric correction and image enhancement processing to obtain a rock sample preprocessing image; S3: Acquisition of rock weathering characteristic parameters: obtaining spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters through rock sample preprocessing images; The spectral characteristic parameters include the wavelength of the reflectivity peak position, the reflectivity ratio of a specific band, and the absorption valley depth; The texture feature parameters include energy, entropy, contrast and LBP features; The morphological characteristic parameters include crack length, crack density and fractal dimension; S4: Construction of rock weathering model: Collect rock multispectral image samples through geological exploration field sample database, laboratory simulated weathering sample database and public rock database, extract spectral, texture and morphological characteristic parameters for each sample image according to the characteristic parameter acquisition process in step S3, combine them into a feature vector, and build a deep convolutional neural network model based on the characteristic parameters; S5: rock weathering degree prediction: input the spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters obtained in step S3 into the trained model to obtain the rock weathering probability vector, and use the maximum probability criterion to judge the probability vector; S6: Interactive feedback: The rock weathering probability vector obtained in step S5, the text labels corresponding to each probability in the rock weathering probability vector, and the judgment result are output to the user end.

[0007] Technical effects and advantages of the present invention: 1. The present invention is more comprehensive and in-depth in feature extraction. In addition to spectral features, texture features are also extracted in detail, such as calculating parameters such as energy, entropy, contrast, etc. through gray-level co-occurrence matrix, and LBP features to describe the complexity of rock surface texture; at the same time, morphological features such as crack length, density and fractal dimension are extracted, making full use of various types of information in multispectral images to more comprehensively characterize rock weathering characteristics; 2. When constructing a deep convolutional neural network model, the present invention uses a large number of samples from a geological exploration field sample database, a laboratory simulated weathering sample database, and a public rock database for training, so that the model can learn rich and diverse rock weathering characteristics, and optimizes the network structure and adjusts the training parameters, such as using an improved VGG16 network structure, reducing the number of neurons in the first two fully connected layers to reduce the complexity of the model, using a stochastic gradient descent optimizer and a suitable learning rate decay strategy, etc., to improve the accuracy and reliability of the model and enhance the generalization ability of the model under complex geological conditions; 3. The present invention can be applied to real-time monitoring of rock weathering. By regularly collecting multi-spectral images of rocks and analyzing them using a trained intelligent detection model, changes in the degree of rock weathering can be discovered in a timely manner. Protective measures can be taken in a timely manner based on the monitoring results to effectively prevent the occurrence of geological disasters, thus meeting the needs for dynamic monitoring of rock weathering in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the process of the present invention.

[0009] Figure 2 It is a schematic diagram of module connection of the present invention. DETAILED DESCRIPTION

[0010] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] As attached Figure 1 A rock weathering intelligent detection method based on multispectral imaging technology is shown, including a multispectral imaging module, an edge computing module, a data preprocessing module, a feature extraction module, a deep learning model module and a user end; In a more specific application of the present invention, the multispectral imaging module includes a Resonon Pika XC multispectral imager, a six-axis robotic arm, a Spectralon diffuse reflection calibration plate, and an ambient light compensation device, which is used to control the imager to perform a 360° surround scan through the robotic arm, obtain a group of 16-band images every 15°, combine the calibration plate and ambient light compensation to ensure image radiation consistency, and output 16-bit TIFF format data with geographic coordinates; The edge computing module is built on the NVIDIA Jetson AGX Orin platform and is used to perform image preprocessing, feature parameter calculation, and model reasoning in real time, supporting data storage and transmission; The data preprocessing module includes radiation correction, geometric correction and image enhancement algorithms. The radiation correction algorithm is used to convert the original grayscale value into absolute reflectance through dark current correction and calibration plate reflectance calculation. The geometric correction algorithm is used to use cubic polynomial transformation and combine 5 feature point matching to achieve image distortion correction. The image enhancement algorithm is used to apply the CLAHE algorithm to improve local contrast. The feature extraction module includes spectral, texture and morphological feature extraction algorithms. The spectral feature algorithm is used to calculate the reflectivity peak position wavelength (760-900nm), 850 / 550nm reflectivity ratio and 2200nm absorption valley depth. The texture feature algorithm is used to extract energy, entropy and contrast based on the gray-level co-occurrence matrix, and to calculate the texture complexity in combination with LBP features. The morphological feature algorithm is used to measure the crack length density through Canny edge detection and Hough transform, and the box dimension method is used to calculate the contour fractal dimension to generate a 23-dimensional feature vector. The deep learning model module is used to input a 23-dimensional feature vector and output a four-dimensional probability value of unweathered, slightly, moderately, and severely weathered; The user end is used to receive the judgment results in real time, display the detection progress, weathering degree thermal map and detailed report; The connection relationship between the above multispectral imaging module, edge computing module, data preprocessing module, feature extraction module, deep learning model module and user end is shown in Figure 2 As shown; The specific implementation of the present invention includes the following steps: S1: Multispectral image acquisition: by inputting preset acquisition parameters into the acquisition terminal, the rock sample image is acquired through the acquisition terminal; It should be specifically stated that the acquisition terminal refers to a multispectral imager with a detector covering the 450-2500nm band, including visible light (450-760nm), near infrared (760-1050nm), and short-wave infrared (1050-2500nm). The spectral resolution of the acquisition terminal is 10nm and the spatial resolution is 0.1mm / pixel; The acquisition parameters refer to dynamically adjusting the exposure time to 50ms through automatic gain control technology to ensure that the image grayscale value is distributed in the range of 100-200, setting the gain to 1.5 times, setting the resolution to 1024×1024 pixels, and adjusting the focal length to 1m to obtain a 1:1 real size image; The method for the acquisition terminal to obtain rock sample images is as follows: A six-axis robotic arm equipped with an imager was used to perform 360° surround acquisition at 1m above the rock surface. A set of images was acquired every 15°, and each set contained 16 bands of data. The exposure time was set to 50ms, the gain was 1.5, and the resolution was 1024×1024 pixels. Twelve images of different perspectives were collected for each rock sample, and each perspective was shot three times in a row to take the average value.

[0012] S2: Multispectral image preprocessing: The collected rock sample images are sequentially subjected to radiation calibration, geometric correction and image enhancement processing to obtain a rock sample preprocessing image; The radiation calibration method is as follows: Absolute radiation calibration is performed using the Spectralon diffuse reflectance calibration plate, and detector noise is eliminated by dark current correction. The calculation formula is:

[0013] Among them, R(λ) is the reflectivity after correction, I(λ) is the gray value of the original image, and I dark (λ) is the dark current image, I cal (λ) is the calibration plate image, R cal (λ) is the standard reflectivity of the calibration plate.

[0014] The geometric correction method is as follows: Using the polynomial transformation method, five obvious feature points in the image are selected as control points and resampled by bilinear interpolation; The image enhancement method is as follows: Use contrast-limited adaptive histogram equalization algorithm, set cliplimit to 40, tilegridsize to 8×8, to enhance the local contrast of the image; S3: Acquisition of rock weathering characteristic parameters: obtaining spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters through rock sample preprocessing images; The spectral characteristic parameters include the wavelength of the reflectivity peak position, the reflectivity ratio of a specific band, and the absorption valley depth; The texture feature parameters include energy, entropy, contrast and LBP features; The morphological characteristic parameters include crack length, crack density and fractal dimension; The calculation of the wavelength of the reflectivity peak position is as follows: for each ROI area of ​​100×100 pixels, the average reflectivity value of each pixel in the area in each band is calculated, and a curve of the reflectivity changing with the band is drawn to find the wavelength corresponding to the peak position. For example, for granite, the wavelength corresponding to the maximum reflectivity is searched in the near-infrared band of 760-900nm. Five different areas are selected for repeated calculation each time the measurement is performed, and the average value is taken to improve data accuracy.

[0015] The specific band reflectivity ratio calculation method is as follows: Calculate the reflectivity ratio of 850nm and 550nm bands, the formula is: For each sample, eight 20 × 20 pixel regions were randomly selected and the average value was calculated.

[0016] The absorption valley depth calculation method is as follows: For the 2200nm absorption valley, the formula for calculating the valley depth is: Ten areas were measured for each sample, and the standard deviation was less than 0.05, which was considered valid data. Taking the absorption valley near 2200nm of clay rock with high montmorillonite content as an example, the data corresponding to 2200nm and its adjacent bands (2150-2250nm) were found in the multispectral image data. For the selected rock area, the reflectivity values ​​of these bands were calculated, and the curve was drawn. The depth of the absorption valley was obtained by calculating the difference between the reflectivity value at the bottom of the absorption valley and the average reflectivity value of the relatively flat areas on both sides of the valley.

[0017] The energy is calculated as follows:

[0018] Where p(i,j) is the element value of gray value i and j in the gray level co-occurrence matrix, which represents the probability of the occurrence of pixel pairs with gray value i and j under a specific spatial relationship, and L is the gray level of the image; The entropy is calculated as follows:

[0019] Where p(i,j) is the element value of gray value i and j in the gray level co-occurrence matrix, which represents the probability of the occurrence of pixel pairs with gray value i and j under a specific spatial relationship, and L is the gray level of the image; The contrast ratio is calculated as follows:

[0020] Where p(i,j) is the element value of grayscale values ​​i and j in the grayscale co-occurrence matrix, representing the probability of the occurrence of pixel pairs with grayscale values ​​i and j under a specific spatial relationship, L is the grayscale level of the image, and n is the absolute difference between grayscale values ​​i and j; It needs to be further explained that Refers to the sum of the probabilities of all pixel pairs (i, j) with a grayscale difference of n. The contrast calculation formula is to weight each grayscale difference n and its corresponding probability sum, with the weight n 2 .

[0021] The calculation method of the LBP feature is as follows: perform LBP transformation on the rock image, use the central pixel value as the threshold to compare with the 8 pixel values ​​in the neighborhood, record the neighborhood pixel value greater than or equal to the central pixel value as 1, otherwise record it as 0, obtain an 8-bit binary number and convert it into a decimal number as the LBP value of the central pixel, traverse the entire image to obtain the LBP value of each pixel, generate an LBP image, count the frequency of different LBP values ​​in the LBP image, construct an LBP histogram, and count the frequency proportion of the high-frequency part in the LBP histogram, that is, the part greater than the average LBP value in the image, as a characteristic parameter reflecting the change in the complexity of the rock surface texture.

[0022] The calculation method of the crack length is as follows: detect the cracks through Canny edge detection and Hough transform, calculate the pixel length of each crack, and convert the pixel length into physical length, that is, 1 pixel = 0.1 mm, and calculate the average crack length.

[0023] The calculation method of the crack density is as follows:

[0024] Where D is the crack density L i is the length of a single crack, A is the image area (unit: mm²); The calculation method of the fractal dimension is as follows: The box dimension method is used to calculate the fractal dimension of the rock contour. First, the rock contour image is binarized, and only the contour part is retained. The contour is covered with square boxes of different sizes (with side lengths of 1 pixel, 2 pixels, 4 pixels, etc.), and the number of boxes N(ϵ) that can just cover the contour under each box size is recorded, where ϵ is the side length of the box. By performing linear fitting on logN(ϵ) and log(1 / ϵ), a goodness of fit R²>0.9 is considered a valid result, and the absolute value of the linear fitting slope is the fractal dimension.

[0025] S4: Construction of rock weathering model: Collect rock multispectral image samples through geological exploration field sample database, laboratory simulated weathering sample database and public rock database, extract spectral, texture and morphological characteristic parameters for each sample image according to the characteristic parameter acquisition process in step S3, combine them into a feature vector, and build a deep convolutional neural network model based on the characteristic parameters; It should be further explained that the geological exploration field sample database is based on typical geological areas in China, such as the North China Craton, the Yangtze Plate, the Qinghai-Tibet Plateau, etc. The laboratory simulated weathering sample database is based on the use of accelerated weathering experimental equipment in the laboratory to simulate physical, chemical and biological weathering processes and generate standard samples. Among them, physical weathering samples are produced by simulating freeze-thaw effects through temperature cycles (-40°C to 80°C), chemical weathering samples are obtained by leaching experiments using 5% HCl solution, and biological weathering samples are obtained by inoculating lichen fungi and culturing for 12 months.

[0026] The rock multispectral image samples are labeled with rock types, and each rock type includes unweathered, slightly weathered, moderately weathered and heavily weathered samples. Each sample has an accurately labeled weathering degree label, and the weathering degree is encoded according to the label. Unweathered is coded as 0, slightly weathered is coded as 1, moderately weathered is coded as 2, and heavily weathered is coded as 3; The steps of constructing the deep convolutional neural network model are as follows: A1: Network structure design: adopt the improved VGG16 network structure, retain the first 5 convolution blocks of the original VGG16, each convolution block contains multiple convolution layers and a pooling layer, reduce the number of neurons in the first two fully connected layers from 4096 to 2048 and 1024 respectively, reduce the model complexity to reduce the risk of overfitting, add a Softmax classification layer at the end of the network, the number of neurons in the Softmax layer is 4, corresponding to the four weathering degrees of rocks (unweathered, slightly weathered, moderately weathered, and heavily weathered), the convolution kernel size in the convolution layer is set to 3×3, the step size is 1, and the padding is 1 to ensure that the size of the feature map remains unchanged after convolution; the pooling layer uses maximum pooling, the pooling kernel size is 2×2, and the step size is 2, which reduces the resolution of the feature map, reduces the amount of calculation and prevents overfitting; A2: Input data processing: Input the sample data after preprocessing and feature extraction into the model, convert each sample feature vector into a tensor form suitable for convolutional neural network input. For example, when the feature vector dimension is 23, expand it to a tensor of size 1×1×23 and perform normalization so that all feature values ​​are in the range of 0-1 to accelerate the convergence of model training; A3: Training parameter setting: A large number of rock multispectral image samples with known weathering degrees were used for training. The initial learning rate was set to 0.001, and the stochastic gradient descent optimizer was used. The momentum parameter was 0.9. During the training process, the learning rate was decayed to 0.1 times the original value every 50 epochs. The training batch size was set to 32. 32 samples were randomly selected from the training set for parameter update each time. A total of 200 epochs were trained. One epoch means a complete training of the entire training set. A4: Data enhancement: To enhance the generalization ability of the model, data enhancement technology is used. The image samples in the training set are randomly rotated ±15°. During each training, 30% of the samples are randomly selected for rotation or flipping to increase the diversity of samples, so that the model can learn a wider range of rock weathering characteristics; A5: Output probability vectors of four weathering degrees; A6: Loss function and optimization: The cross entropy loss function is used to measure the difference between the model prediction result and the true label. The calculation formula is:

[0027] Where N represents the number of samples, 4 is the number of categories, which corresponds to the four weathering degrees of rocks, and y ob is the true label of sample o belonging to category b. If sample o does belong to category b, then y ob The value is 1, otherwise it is 0. ob It is the probability that the model predicts that sample o belongs to category b, and its value is between 0 and 1.

[0028] It should be further explained that during the training process, the optimizer continuously adjusts the model parameters to minimize the loss function value, and the parameters that minimize the loss function are used as the final model parameters.

[0029] Specifically, taking a 23-dimensional feature vector (containing 10 spectra + 8 textures + 5 morphologies) as an example, the process of building a deep convolutional neural network model is as follows: B1: Perform Min-Max normalization on the 23-dimensional feature vector to convert the feature vector into a four-dimensional tensor: X∈R 1 ×1×23 ; B2: Fully connected layer processing: including the first fully connected layer and the second fully connected layer; First fully connected layer: z1=ReLU(W1X+b1), where W1∈R 2048×23 is a weight matrix containing 2048×23=47,024 learnable parameters, b1∈R 2048 is the bias vector, ReLU is the activation function: ReLU(z)=max(0,z); The second fully connected layer: z2=ReLU(W2z1+b2), where W2∈R 1024×2048 is a weight matrix containing 1024×2048=2,097,152 parameters, b2∈R 1024 is the bias vector; B3: Softmax classification layer: including raw score calculation and probability normalization; Original score calculation: z = W3z2 + b3, where W3∈R 4×1024 is the weight matrix, containing 4×1024=4,096 parameters, b3∈R 4 is the bias vector; Probability normalization: , where b=(0,1,2,3); B4: Output layer: The output is a four-dimensional probability vector P=[p0,p1,p2,p3], where p0 refers to the probability of unweathered, p1 refers to the probability of slightly weathered, p2 refers to the probability of moderately weathered, and p3 refers to the probability of severely weathered.

[0030] S5: rock weathering degree prediction: input the spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters obtained in step S3 into the trained model to obtain the rock weathering probability vector, and use the maximum probability criterion to judge the probability vector; It should be specifically noted that the maximum probability criterion specifically refers to: when the maximum probability value p max ≥0.7, directly output p max Corresponding category; when p maxWhen <0.7, the uncertainty processing mechanism is activated and the "pending review" label is output; S6: Interactive feedback: The rock weathering probability vector obtained in step S5, the text labels corresponding to each probability in the rock weathering probability vector, and the judgment result are output to the user end.

[0031] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A rock weathering intelligent detection method based on multispectral imaging technology, characterized in that: include: S1: Multispectral image acquisition: by inputting preset acquisition parameters into the acquisition terminal, the rock sample image is acquired through the acquisition terminal; S2: Multispectral image preprocessing: The collected rock sample images are sequentially subjected to radiation calibration, geometric correction and image enhancement processing to obtain a rock sample preprocessing image; S3: Acquisition of rock weathering characteristic parameters: obtaining spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters through rock sample preprocessing images; The spectral characteristic parameters include the wavelength of the reflectivity peak position, the reflectivity ratio of a specific band, and the absorption valley depth; The texture feature parameters include energy, entropy, contrast and LBP features; The morphological characteristic parameters include crack length, crack density and fractal dimension; S4: Construction of rock weathering model: Collect rock multispectral image samples through geological exploration field sample database, laboratory simulated weathering sample database and public rock database, extract spectral, texture and morphological characteristic parameters for each sample image according to the characteristic parameter acquisition process in step S3, combine them into a feature vector, and build a deep convolutional neural network model based on the characteristic parameters; S5: rock weathering degree prediction: input the spectral characteristic parameters, texture characteristic parameters and morphological characteristic parameters obtained in step S3 into the trained model to obtain the rock weathering probability vector, and use the maximum probability criterion to judge the probability vector; S6: Interactive feedback: The rock weathering probability vector obtained in step S5, the text labels corresponding to each probability in the rock weathering probability vector, and the judgment result are output to the user end.

2. The method for intelligent detection of rock weathering based on multispectral imaging technology according to claim 1 is characterized in that: The radiation calibration method is as follows: Absolute radiation calibration is performed using the Spectralon diffuse reflectance calibration plate, and detector noise is eliminated by dark current correction. The calculation formula is: Among them, R(λ) is the reflectivity after correction, I(λ) is the gray value of the original image, and I dark (λ) is the dark current image, I cal (λ) is the calibration plate image, R cal (λ) is the standard reflectivity of the calibration plate; The geometric correction method is as follows: Using the polynomial transformation method, five obvious feature points in the image are selected as control points and resampled by bilinear interpolation; The image enhancement method is as follows: Use the contrast-limited adaptive histogram equalization algorithm, set the cliplimit to 40, and the tilegridsize to 8×8 to enhance the local contrast of the image.

3. The intelligent rock weathering detection method based on multispectral imaging technology according to claim 1 is characterized in that: The calculation of the wavelength of the reflectivity peak position is as follows: for each ROI area of ​​100×100 pixels, the average reflectivity value of each pixel in the area in each band is calculated, a curve of reflectivity changing with band is drawn, and the wavelength corresponding to the peak position is found; The specific band reflectivity ratio calculation method is as follows: Calculate the reflectivity ratio of 850nm and 550nm bands, the formula is: For each sample, eight 20 × 20 pixel regions were randomly selected and the average value was calculated; The absorption valley depth calculation method is as follows: For the 2200nm absorption valley, the formula for calculating the valley depth is: The energy is calculated as follows: Where p(i,j) is the element value of gray value i and j in the gray level co-occurrence matrix, which represents the probability of the occurrence of pixel pairs with gray value i and j under a specific spatial relationship, and L is the gray level of the image; The entropy is calculated as follows: Where p(i,j) is the element value of gray value i and j in the gray level co-occurrence matrix, which represents the probability of the occurrence of pixel pairs with gray value i and j under a specific spatial relationship, and L is the gray level of the image; The contrast ratio is calculated as follows: Where p(i,j) is the element value of grayscale values ​​i and j in the grayscale co-occurrence matrix, representing the probability of the occurrence of pixel pairs with grayscale values ​​i and j under a specific spatial relationship, L is the grayscale level of the image, and n is the absolute difference between grayscale values ​​i and j; The calculation method of the LBP feature is as follows: perform LBP transformation on the rock image, compare the central pixel value with 8 pixel values ​​in the neighborhood as a threshold, record the neighborhood pixel value greater than or equal to the central pixel value as 1, otherwise record it as 0, obtain an 8-bit binary number and convert it into a decimal number as the LBP value of the central pixel, traverse the entire image to obtain the LBP value of each pixel, generate an LBP image, count the frequency of different LBP values ​​in the LBP image, construct an LBP histogram, and count the frequency proportion of the high-frequency part in the LBP histogram, that is, the part greater than the average value of the LBP value in the image, as a characteristic parameter reflecting the change in the complexity of the rock surface texture; The crack length is calculated as follows: cracks are detected by Canny edge detection and Hough transform, the pixel length of each crack is calculated, and the pixel length is converted into a physical length, that is, 1 pixel = 0.1 mm, and the average crack length is calculated; The calculation method of the crack density is as follows: Where D is the crack density L i is the length of a single crack, and A is the image area.

4. The intelligent detection method for rock weathering based on multispectral imaging technology according to claim 1 is characterized in that: The steps of constructing the deep convolutional neural network model are as follows: A1: Network structure design: adopt the improved VGG16 network structure, retain the first 5 convolution blocks of the original VGG16, each convolution block contains multiple convolution layers and a pooling layer, reduce the number of neurons in the first two fully connected layers from 4096 to 2048 and 1024 respectively, reduce the complexity of the model, add a Softmax classification layer at the end of the network, the number of neurons in the Softmax layer is 4, corresponding to the four weathering degrees of the rock, the convolution kernel size in the convolution layer is set to 3×3, the step size is 1, and the padding is 1. The pooling layer uses maximum pooling, the pooling kernel size is 2×2, and the step size is 2; A2: Input data processing: input the sample data after preprocessing and feature extraction into the model, and convert each sample feature vector into a tensor form; A3: Training parameter setting: Use rock multispectral image samples with known weathering degree for training, set the initial learning rate to 0.001, use stochastic gradient descent optimizer, momentum parameter to 0.9, and the learning rate decays to 0.1 times of the original every 50 epochs during training. The training batch size is set to 32, and 32 samples are randomly selected from the training set for parameter update each time. A total of 200 epochs are trained, and one epoch means a complete training of the entire training set. A4: Randomly rotate the image samples in the training set by ±15°. During each training, randomly select 30% of the samples for rotation or flipping. A5: Output probability vectors of four weathering degrees; A6: Loss function and optimization: The cross entropy loss function is used to measure the difference between the model prediction result and the true label. The calculation formula is: Where N represents the number of samples, 4 is the number of categories, which corresponds to the four weathering degrees of rocks, and y ob is the true label of sample o belonging to category b. If sample o does belong to category b, then y ob The value is 1, otherwise it is 0. ob It is the probability that the model predicts that sample o belongs to category b, and its value is between 0 and 1.

5. The intelligent rock weathering detection method based on multispectral imaging technology according to claim 1 is characterized in that: The maximum probability criterion specifically refers to: when the maximum probability value p max ≥0.7, directly output p max Corresponding category; when p max When <0.7, the uncertainty processing mechanism is activated and the "pending review" label is output.

Images