Introduction

Acidification of groundwater is increasingly affected by environmental pollution1,2,3,4,5. Continuous erosion of rock mineral components by acidic groundwater will lead to geological hazards such as ground subsidence and landslides6,7,8,9,10, especially in areas with fragile topography or consisting of dissolved rocks, causing it difficult to carry out underground engineering safely11,12,13. The mineral composition and the chemical weathering process of the rock is accelerated by the acidic groundwater, and the internal structure of the rock is changed. Due to the dissolution of the minerals, the original space occupied by the minerals becomes empty, increasing the porosity of rocks. It is reduces the density and strength of rocks, changing the permeability and affecting the flow and distribution of groundwater on the rocks14,15,16. In response to the potential risks posed by acidic groundwater to geotechnical engineering, the urgent need for geological assessment by means of sound monitoring is essential.

Acoustic emission17,18,19,20, CT21,22, infrared thermal imaging23,24,25 and other technologies have been widely used to monitor rock damage. However, the signal of mineral dissolution caused by acidic groundwater is lower frequency than that of rock fracture, and acoustic emission technology is difficult to distinguish. The CT technology can not achieve in-situ, long-term monitoring, it is difficult to meet the stage evaluation of rock acid corrosion process. Infrared thermal imaging technology can only capture the abnormal changes on the rock surface, and it is difficult to reveal the internal damage caused by acid corrosion. Mid-infrared spectroscopy is an effective detection method for real-time and non-destructive monitoring of geological hazards26, which can reflect the characteristics of rock mineral based on the differences in the reflective positions and intensities generated by different groups. Many scholars have carried out a lot of research in the field of rock mechanics using infrared spectroscopy27,28,29. For example, Li et al.30 obtained more abundant rock spectral information and classified the lithology based on the hyperspectral technology of continuous band. He et al.31 used the near-infrared spectroscopy technology to test the spectral characteristics of water-bearing conglomerate, siltstone and compacted soil, respectively. The experimental results showed that peak height, right shoulder width and peak area had the highest correlation with the spectral characteristics of water-bearing rocks. Huang et al.32 investigated the mid-infrared spectral features of sandstone, granite and marble during loading in outdoor, and the results showed that the spectral features changed most significantly during the fracturing stage. Wu et al.33 classified hyperspectral images of 81 common magmatic and metamorphic rocks, and divided the rocks into 9 major and 28 minor categories. Liu et al.34 explored the spectral reflectance of lignite, bituminous coal and anthracite. The results of the study showed that the granularity of 0.1 mm is the sensitive limit of the spectral features, and the change of reflectance spectral features of the coal samples with the granularity is not obvious for greater than 0.1 mm. When the smaller the granularity is, the greater the slope of the spectral reflectance curve is of the coal samples for less than 0.1 mm. Ma et al.35 used infrared thermal imaging to investigate the mudstone samples, which were subjected to chemical corrosion using both strong acid and neutral solutions. The stress-strain damage constitutive model based on average infrared temperature was established. Shen et al.36 studied the concrete-rock composite under acid attack and preload conditions (CRC), and found that the peak strength and elastic modulus of CRC samples treated with acid solution and preloaded were reduced. The above studied have shown that the spectral technique is widely used in coal and rock materials, which can be used as one of the reliable detection techniques in the field of geotechnical engineering. However, the correspondence between the spectral characteristics and acidification degrees of rock samples remains worthy to further explore. Therefore, it is necessary and meaningful to determine the mid-infrared band spectral features of rocks with different acidification degrees, reflecting the mechanism of weakening the mechanical strength of rocks.

In this study, red sandstone, a calcium carbonate-rich rock, was subjected to immersion in hydrochloric acid solutions of varying concentrations over different durations. The objective was to monitor the spectral characteristics of the rock samples under different acidification conditions. By integrating RF, KNN, and SVM algorithms, a predictive model for assessing the acidification degree of rock was developed. This study provides a scientific foundation and an effective methodology for evaluating the acidification effects pertinent to geotechnical engineering applications.

Experimental setup

Sample Preparation

The red sandstone samples were all selected from Wuding, Chuxiong, Yunnan Province, China. In order to avoid the interference of moisture on the spectral intensity during the processing, the samples were placed in a constant-temperature drying oven at 105 °C for 24 h before the beginning of the experiment, until the dry weight was constant. At the same time, to clarify the micro-mineral composition and the degree of mineral crystallisation inside the sample, the main mineral composition of the red sandstone was measured as detrital minerals and clay minerals by X-ray diffraction experiments, as shown in Fig. 1.

Fig. 1
figure 1

Schematic diagram of mineral composition on red sandstone.

Full size image

Acquisition of spectral curves

The Foil 20-Z Fourier Transform Mid-Infrared Spectroscopy system was utilized to acquire spectra from red sandstone samples within the wavelength range of 2.5 to 25 μm, with a data sampling interval of 0.5 μm and a probe resolution of 4 cm. Spectral signals were collected via diffuse reflection, as illustrated in Fig. 2. In a controlled environment maintained at constant temperature and humidity, mid-infrared spectra of samples subjected to varying acidification times and concentrations were measured, obtaining 1600 mid-infrared spectral curves. The detailed experimental procedure is described as follows:

  1. (1)

    The red sandstone samples that were not acidified after drying were taken out, cooled to room temperature, the surface dust was removed and fixed at the test position of the mid-infrared spectrometer. The angle between the spectral probe and sample was adjusted to 90°, and the mid-infrared spectral curve with the acidification degree of 0% was measured.

  2. (2)

    After drying, the unacidified red sandstone samples were placed in hydrochloric acid solution with pH of 1, 2, 3 and 4, and soaked for 1 h, 3 h, 5 h, 24 h and 72 h, respectively. The pH value of the solution was detected with test paper every 30 min to ensure that the pH of the hydrochloric acid solution was constant in the experiment.

  3. (3)

    The samples soaked in hydrochloric acid solution with pH of 1, 2, 3 and 4 for 72 h were taken out and placed in a electronic balance to observe the mass. When the dry weight was constant, the free water on the surface of the sample was wiped and the mid-infrared spectral curve of 100% was measured.

  4. (4)

    The samples soaked in hydrochloric acid solution at different pH for 1 h, 3 h, 5 h and 24 h were taken out, and the mid-infrared spectral curves of samples at different acidification times and pH values were measured.

  5. (5)

    The collected spectral curves were processed by S-G convolution smoothing to achieve the preliminary filtering and denoising effect, and comparative analysis is carried out.

Fig. 2
figure 2

The flow chart of mid-infrared spectrum experiments.

Full size image

Middle infrared spectral characteristics of samples

Physical characteristic analysis

The rocks are natural geological materials cemented by a variety of mineral crystal particles. The main components of red sandstone minerals in the natural state include quartz (SiO2), calcite (CaCO3), potassium feldspar (KAlSi3O8), plagioclase ((Na, Ca)AlSi3O8), hematite (Fe2O3), etc. It contains calcite and various feldspar minerals accounting for up to 95% of the total mineral composition. Therefore, the red sandstone is easily soluble in acidic solutions, and the chemical reactions between calcite, feldspar minerals and hydrochloric acid occur as follows:

$${\text{CaC}}{{\text{O}}_3}+2{\text{HCl}} \to {\text{CaC}}{{\text{l}}_2}+{{\text{H}}_2}{\text{O}}+{\text{C}}{{\text{O}}_2} \uparrow {\text{~}}$$
(1)
$${\text{KAlS}}{{\text{i}}_3}{{\text{O}}_8}+4{\text{HCl}}+{{\text{H}}_2}{\text{O}} \to {\text{KCl}}+{\text{AlC}}{{\text{l}}_3}+3{{\text{H}}_2}{\text{Si}}{{\text{O}}_3} \downarrow$$
(2)
$${\text{NaAlS}}{{\text{i}}_3}{{\text{O}}_8}+4{\text{HCl}}+{{\text{H}}_2}{\text{O}} \to {\text{NaCl}}+{\text{AlC}}{{\text{l}}_3}+3{{\text{H}}_2}{\text{Si}}{{\text{O}}_3} \downarrow$$
(3)
$${\text{CaAlS}}{{\text{i}}_3}{{\text{O}}_8}+4{\text{HCl}}+{{\text{H}}_2}{\text{O}} \to {\text{CaCl}}+{\text{AlC}}{{\text{l}}_3}+3{{\text{H}}_2}{\text{Si}}{{\text{O}}_3} \downarrow$$
(4)

.

A variety of mineral components that bearing the strength of the rock sample in an acidic solution are fused with the acid solution. A large number of calcite and feldspar minerals in the sample react chemically with \(\:{\text{H}}^{+}\) and \(\:{\text{C}\text{l}}^{-}\) in hydrochloric acid solution. With the increase of acid solution concentration and soaking time, the internal pores of the rock sample gradually expanded with the intensity of chemical reaction, exhibiting deterioration of the mechanical properties of the sample. The mid-infrared Fourier transform spectrometer is used to detect the spectral intensity of acidified samples and compare it with unacidified samples determining the degree of acidification of samples, which provides a strong guarantee for the safe development of geotechnical underground engineering.

Spectral characteristic analysis

Soaking in different pH solutions

Therefore, the soaking time and solution concentration are regarded as two characteristic variables, and the mid-infrared spectral curves of red sandstone samples under different boundary conditions are collected for comparative analysis. Firstly, the mid-infrared spectral curve of red sandstone samples soaked in hydrochloric acid solution at different pH values for the same time was collected, as shown in Fig. 3. As can be seen from the Fig. 3, the spectral intensity increases significantly with the increase of pH value of hydrochloric acid solution. Due to the hydrochloric acid solution with a higher concentration contains more abundant \(\:{\text{H}}^{+}\) and \(\:{\text{C}\text{l}}^{-}\), and the chemical reaction with various mineral components in the red sandstone sample is more intense. With the progress of the reaction, the crystal lattice structure of the mineral is destroyed, resulting in changes in the physical properties of the rock, the more metal cations precipitated, such as \(\:{\text{C}\text{a}}^{+},\:{\text{K}}^{+},\:{\text{N}\text{a}}^{+}\), the higher the dissolution degree of the rock, the less mineral composition of the sample itself, and the lower the reflected spectral intensity.

Fig. 3
figure 3

Mid-infrared spectral curves of red sandstone samples in hydrochloric acid solutions at different pH.

Full size image

Different soaking duration

Fig. 4
figure 4

shows the mid-infrared spectral curve of red sandstone samples soaked in acid solution for different times. It can be seen that with the increase of soaking time, the rock surface becomes rougher or porous, more sample minerals are dissolved at each concentration of hydrochloric acid, and the characteristic peaks of carbonate ions weaken or disappear, resulting in a decrease in spectral reflection intensity.

Full size image

Spectral data processing method and model

Spectral pretreatment

In order to efficiently extract the important feature from a large number of spectral data, the original spectral data is subtracted from the mean value. Then, the expected standard deviation of the spectral data after the difference is calculated by the quotient, and finally the standardized spectral data is obtained. The calculation formula is as follows:

$$A=\frac{{a - mean\left( a \right)}}{{std\left( a \right)}}$$
(5)

.

where a is the original spectral data, \(\:mean\left(a\right)\) is the average value of the original spectral data, \(\:std\left(a\right)\) is the standard deviation of the original spectral data, and \(\:A\) is the standardized spectral data.

Subsequently, the PCA method is used for data processing. The PCA is a data dimensionality reduction method based on continuous attributes, which can screen out representative characteristic variables from the miscellaneous spectral data to achieve the compression function. Through the orthogonal transformation of the original spectral data to remove the correlation of the original spatial base data, the principal component and its contribution rate are obtained, and the maximum extent of spectral data reduction and noise filtering is achieved.

If that the observed data matrix of the original spectral variable for n times is:

$$\delta =\left( {\begin{array}{*{20}{c}} {\begin{array}{*{20}{c}} {{\delta _{11}}}&{{\delta _{12}}} \\ {{\delta _{21}}}&{{\delta _{22}}} \end{array}}& \cdots &{\begin{array}{*{20}{c}} {{\delta _{1p}}} \\ {{\delta _{2p}}} \end{array}} \\ \vdots & \ddots & \vdots \\ {\begin{array}{*{20}{c}} {{\delta _{n1}}}&{{\delta _{n2}}} \end{array}}& \cdots &{{\delta _{np}}} \end{array}} \right)=\left( {{\delta _1},{\delta _2}, \cdots ,{\delta _p}} \right)$$
(6)

.

To calculate its correlation coefficient matrix \(\:R={\left({r}_{ij}\right)}_{p\times\:p}\), \(\:{r}_{ij}\) can be expressed as:

$${r_{ij}}=\mathop \sum \limits_{{k=1}}^{n} \left( {{\delta _{ki}} - \overline {{{\delta _i}}} } \right)\left( {{\delta _{kj}} - \overline {{{\delta _j}}} } \right)/\sqrt {\mathop \sum \limits_{{k=1}}^{n} {{\left( {{\delta _{ki}} - \overline {{{\delta _i}}} } \right)}^2}\mathop \sum \limits_{{k=1}}^{n} {{\left( {{\delta _{kj}} - \overline {{{\delta _j}}} } \right)}^2}} ~~({r_{ij}}={r_{ji}},~{r_{ii}}=1)$$
(7)

.

The characteristic root of the correlation coefficient matrix \(\:\eta\:\) is obtained, determining the number of principal components m, and calculate the contribution rate of principal components is:

$$\:{Z}_{1}={\beta\:}_{1i}{\delta\:}_{1}+{\beta\:}_{2i}{\delta\:}_{2}+\cdots\:+{\beta\:}_{pi}{\delta\:}_{p}\:\:\:\:\:(i=\text{1,2},\cdots\:,m)\:\:$$
(8)

.

where \(\:\beta\:\) is the corresponding unit eigenvector of each principal component.

The spectral data of red sandstone samples are standardized and analyzed by principal component analysis. It can be seen from Fig. 5, when the first three principal components are selected, the cumulative contribution rate has reached 95%. It shows that the first two principal components can be extracted to calculate the model that can predict the different acidification degree of red sandstone.

Fig. 5
figure 5

Results of PCA principal component analysis.

Full size image

Model principle and method

KNN algorithm

The K-Nearest Neighbors (KNN) algorithm can determine the category and obtain predicted values via measuring the distance between random and all samples in the training set. It is worth noting that KNN algorithm does not have a traditional training process. Moreover, the distance between a test sample and all training samples are calculated, and the nearest K training samples are selected as the nearest neighbors of the test samples. Then, the category of the test sample is determined by the majority voting method, that is, the category with the largest number of K neighbors is extracted as the result. The Euclidean distance d is calculated as follows:

$$d=\sqrt {\mathop \sum \limits_{{i=1}}^{n} {{\left( {{x_i} - {y_i}} \right)}^2}}$$
(9)

.

where \(\:{x}_{i}\) and \(\:{y}_{i}\) are the values of test samples and training samples on the i dimensional feature, respectively, and n is the number of test samples.

Due to the high computational complexity of the KNN algorithm, it is more sensitive to the scale and noise of the data. Hence, the S-G filtering of the data are required to adopt before modeling. Then, the pre-processing of standardization and PCA algorithm are adopted, and then Euclianian distance is used to find out the K samples with the nearest distance. The K was chosen as 2, and the principle of KNN algorithm is shown in Fig. 6.

Fig. 6
figure 6

Schematic diagram of KNN algorithm.

Full size image

SVM algorithm

Due to the KNN algorithm is more suitable for small-scale and low-dimensional data sets, it has certain requirements for feature scaling and noise processing. Therefore, the Support Vector Machine (SVM) algorithm is selected to model the spectral data. The fundamental principle of the SVM algorithm is to identify an optimal hyperplane that maximally separates different classes of data points while maintaining the greatest possible margin between them and avoiding all sample points. For non-linearly separable data, SVM achieves linear separation by mapping the data into a higher-dimensional space through the introduction of a kernel function. The Gaussian (RBF) kernel function was selected, and its calculation formula presented as follows:

$$K\left( {x,x^{\prime}} \right)=exp\left( { - \gamma x - {{x^{\prime}}^2}} \right)$$
(10)

.

where, \(\:\gamma\:\) is the width of Gaussian distribution. The larger the value, the smaller the influence range of data points and the more complex the model. The smaller the value, the larger the range of influence and the smoother the model. To reduce the training errors, the penalty factor was set to 10, which has a high penalty for misclassification. And the Gamma value, which controls the range of influence of the RBF kernel, was set to 1 for moderate smoothing of the spectral data. The SVM algorithm is shown in Fig. 7.

Fig. 7
figure 7

Schematic diagram of SVM algorithm.

Full size image

RF algorithm

Due to the complexity in optimizing SVM algorithm parameters, to select the most appropriate prediction model for the spectral acidification of rock samples, the Random Forest (RF) method is employed. This approach constructs multiple decision trees and integrates their prediction outcomes to enhance both the accuracy and stability of the model. Specifically, RF utilizes bootstrap resampling technology to repeatedly and randomly select several samples from the original spectral training dataset to form new training sets. Subsequently, it generates numerous classification trees via resampling method to constitute a random forest. The random forest classifier can be used to classify the spectral test set, and the classification results are summarized, as shown in Fig. 8. For modeling the spectral data of various acidizing processes, the parameters were configured such that the number of trees in the forest was set to 200, the minimum sample size required to split an internal node was set to 1, and the minimum information gain required to split a node was set to 0 prior to classification.

Fig. 8
figure 8

Schematic diagram of RF algorithm.

Full size image

In order to further optimize the accuracy of automatic prediction of acidification spectral features of rock samples, a fusion model was established by combining KNN, SVM and RF algorithm with voting mechanism. The classification results of different models for rock sample spectral data were voted, and the category with the most frequency was selected as the final classification result to build a more powerful and robust intelligent prediction model.

Model evaluation

In order to compare and illustrate the advantages and disadvantages of the model, the determination coefficient R2, mean square error (MSE), root mean square error (RMSE) and cross validation root mean square error (RMSECV) are selected to verify the model. The R2 value of the determination coefficient represents the degree of correlation between the predicted value and the actual value. The closer the value is to 1, the closer the two are, indicating the better the prediction effect of the model. The calculation formula is written as follows:

$${R^2}=1 - \frac{{\mathop \sum \nolimits_{{i=1}}^{n} {{\left( {{y_i} - \overline {{{y_i}}} } \right)}^2}}}{{\mathop \sum \nolimits_{{i=1}}^{m} {{\left( {{y_i} - \bar {y}} \right)}^2}}}$$
(11)

.

where, n is the number of the samples, the \(\:{y}_{i}\) is true value, the \(\:\stackrel{-}{y}\) is refer to sample mean, and \(\:\stackrel{-}{{y}_{i}}\) is the sample forecast mean.

The MSE is more accurate in evaluating the model and can reflect the instability of the model to a certain extent. Its calculation formula is expressed as follows:

$$MSE=\frac{1}{n}\mathop \sum \limits_{{i=1}}^{n} {\left( {{{\hat {y}}_l} - {y_i}} \right)^2}$$
(12)

.

where, \(\:\widehat{{\text{y}}_{\text{i}}}\) is the predicted value. If the difference between the predicted value and the true value is large, the MSE value is larger.

$$\:RMSE=\sqrt{\frac{1}{n}\sum\:_{i=1}^{n}{\left(\widehat{{y}_{i}}-{y}_{i}\right)}^{2}}$$
(13)

.

The RMSE can more intuitively represent the error between the predicted value and the real value. The larger the model error, the larger the RMSE value.

The MAE is a measure of the mean absolute difference between the predicted value and the true value. The calculation formula is written as follows:

$$MAE=\frac{1}{n}\mathop \sum \limits_{{i=1}}^{n} \left| {{y_i} - {{\hat {y}}_i}} \right|$$
(14)

.

The MAE reflects the average deviation between the predicted value and the actual value. And the smaller the MAE, the more accurate the model prediction.

Spectral prediction model and evaluation of acidified samples

The spectral intensity of rock samples soaked for 72 h at various pH concentrations, which exhibited the highest degree of acidification, was designated as level 1. The remaining samples with different soaking times were categorized into varying degrees of acidification based on the spectral intensity. The spectral data of different acidification levels were partitioned in a 3:1 ratio. The spectral data often exhibits local similarity in feature space, e.g., adjacent wavelengths correlate with similar mineral compositions. The non-parametric nature of KNN effectively captures these local patterns without assuming global data distributions. The dynamic weight allocation based on the acid concentration gradients are introduced, which is help to enhance the adaptability of KNN to nonlinear spectral shifts caused by mineral dissolution. The samples with similar characteristics were classified together by using KNN model, which is resulting in the development of predictive models for the acidification degrees of rock samples soaked at different pH values and durations, as illustrated in Fig. 9.

Fig. 9
figure 9

Comparison of acidification degree predicted and real by KNN model.

Full size image

As can be seen from Fig. 9, the acidification degree on the surface of test samples is relatively high, which is due to the high concentration of hydrochloric acid solution and strong erosion of rock samples. With the dilution of hydrochloric acid concentration, the acidification degree on the surface and inside of samples decreases. Table 1 shows the evaluation results of the KNN model. It can be seen from that the coefficient of determination R2, mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) all show good prediction effect under different pH values.

Table 1 KNN model evaluation results.
Full size table

The high spectral data volume and dimension need to be processed. The kernel technique of SVM maps data into a separable space while minimizing over-fitting. Figure 10 presents the prediction results of SVM model for rock samples with different degrees of acidification. It can be seen from that the results of SVM model are not as ideal as those of KNN model, and there are errors in several test samples under different hydrochloric acid concentrations, causing unsatisfactory model evaluation results.

Fig. 10
figure 10

Comparison of acidification degree predicted and real by SVM model.

Full size image

As can be seen from Table 2, for the prediction model of acidification degree of rock samples soaked at different pH values for different times, the overall prediction effect of SVM model is general, and the model accuracy is low.

Table 2 SVM model evaluation results.
Full size table

The spectral noise, e.g., instrument drift, scattering effects requires a set approach to improve robustness. The feature bagging of RF and majority voting essentially help suppress noise interference. Integrated spectral band importance ranking guides tree segmentation, focusing on acid-sensitive wavelength ranges. In order to establish the best prediction model for acidification degree of rock samples, the random forest model was selected for comparative analysis. The comparison results between the predicted degree of acidification degree and the real degree of acidification degree by RF model were shown in Fig. 11. As can be seen from the figure, under different hydrochloric acid concentration conditions, the corresponding relationship between the test value and the predicted value is more consistent than that of the SVM model and the KNN model, and the prediction effect is better.

Fig. 11
figure 11

Comparison of acidification degree predicted and real by RF model.

Full size image

Table 3 shows the evaluation results of RF model on acidification degree of rock samples. It can be seen from the table that the R2 coefficients of random forest model are all above 0.9, indicating that the model has better prediction results and is more suitable for predicting acidification degree of rock samples than SVM model and KNN model. In order to further improve the reliability and accuracy of the prediction model for rock acidification degree, the KNN, SVM and RF random forest models mentioned above were selected to combine to predict rock samples with different durations of soaking under different hydrochloric acid concentrations. To combine the boundary-sensitive hyperplanes of SVM, the band importance of RF, and local similarity of KNN metrics into a unified feature space. The results with more votes were selected as the prediction results for rock acidification degree. The confusion matrix is used to show the correct and wrong prediction of the fusion model. Among them, confusion matrix is a classification error matrix based on the number of sample categories, and the number of elements on the main diagonal determines the correct classification situation, so as to evaluate the merits of the model.

Table 3 RF model evaluation results.
Full size table
Fig. 12
figure 12

The result of the confusion matrix of the fusion model.

Full size image

Figure 12 shows the classification results of whether the prediction results are correct or not by the confusion matrix of rock samples with different pH values. The accuracy of the fusion model is as high as 0.97, 0.94, 0.95 and 0.94 respectively, and the accuracy of the comprehensive model is 0.95. Compared with KNN model, SVM model and RF model, the prediction results of the fusion model are the most reliable.

Conclusions

In this study, the red sandstone samples were placed under different hydrochloric acid concentration and soaking time. To combine with the mid-infrared spectrum characteristics, the response relationship between spectral intensity and acidification degree in the acidification process of red sandstone was analyzed, and a prediction model of acidification degree of red sandstone based on machine learning was established. The main conclusions are as follows:

  1. (1)

    The different concentrations of hydrochloric acid solutions result in varying degrees of mineral particle dissolution within the rock samples, leading to differences in infrared reflectance spectral intensity. Specifically, higher concentrations and longer soaking times lead to the weakening or disappearance of carbonate ion characteristic peaks, thereby reducing spectral reflectance intensity and increasing the degree of acidification.

  2. (2)

    The SG convolutional smoothing, standardization and PCA pretreatment methods were used to reduce the dimensionality of the original acidified samples. Then, the KNN, SVM and RF models were applied to predict the acidification degree of the red sandstone samples, in which the order of prediction results was RF > KNN > SVM. In order to improve the accuracy of acidification detection model, a fusion model was proposed, which further integrated the advantages of KNN, SVM and RF models. The results show that the average accuracy of red sandstone samples with different acidification degrees reaches 95%, which is significantly improved compared with the single model.

  3. (3)

    A non-destructive in-situ assessment of acidification degree of red sandstone is realized, which can be integrated into the geological engineering monitoring system and provide key technical support for real-time risk early warning of tunnels and slopes in acid rain erosion areas.