1 Introduction

Geomechanical characterization of rock masses is important for various applications, such as the preliminary design of civil engineering works, the assessment of geological risks, assessment of the quality of quarry materials for the construction and paving industry, and more, and is based on both laboratory and on-site investigations. The rock mass characterization process includes the determination of structure geometries, the spatial distribution and condition of rock discontinuities, as well as the mechanical properties of the intact rock (e.g. Bakun-Mazor et al. 2009; Hatzor et al. 2010). However, there are significant pitfalls in the conventional approaches for the characterization of rock masses: (1) there is often limited access to the outcrop; (2) extracting material from the field for laboratory testing may be subject to severe sampling biases; and (3) both in situ and laboratory tests are costly and time-consuming.

1.1 Remote Sensing in Engineering Geology

In view of the limitations of the conventional approaches, remote sensing (RS) technologies have increasingly been adopted during the past two decades, enabling data acquisition and geological mapping of regions without any physical contact with the areas being explored. Advances in RS technologies have led to the development of new methods for geotechnical engineers and geologists when mapping surface geometry and rock discontinuities; many publications in recent years have demonstrated the capabilities of LiDAR and photogrammetry for rock mass evaluation and characterization. These include discontinuity mapping (Kemeny and Post 2003; Umili et al. 2013; Vöge et al. 2013), rock-fall modeling along transportation corridors (Feng and Röshoff 2004; Kemeny 2008; Lato et al. 2009, 2012), and kinematic evaluation in mountainous terrain (Collins and Sitar 2004; Haneberg 2008; Sturzenegger and Stead 2009; Zimmer et al. 2012; Curtaz et al. 2014; Cheng et al. 2021). These examples, and many others, employ RS technologies to produce three-dimensional (3D) point cloud models that enable the evaluation of discontinuity characteristics such as orientation, spacing/intensity, and persistence, to assess the potential instability of rock masses (Lato and Vöge 2012). An overview of unmanned aerial vehicles and their potentiality in several engineering geology applications was recently presented (Giordan et al. 2020). While information regarding the spatial distribution of discontinuities can already be obtained by means of RS techniques, the mechanical characteristics of the intact rock cannot yet be assessed by these techniques. In this paper, we utilize reflectance spectroscopy at high spectral and spatial resolution as a basis for hyperspectral RS (HRS) technology to evaluate the mechanical properties of the intact rock material.

1.2 Hyperspectral Remote Sensing of Minerals

RS examines the interaction of electromagnetic radiation (EMR) with matter that results from the chemical and physical characteristics of the sensed matter. In general, we measure the reflected or emitted EMR in the 0.3 μm to 1 m (m) wavelength range. Earth RS scientists categorize this region into several ranges based on the physical interaction phenomena. The visible to near infrared (VNIR, 0.4—1.0 μm) range is reflected EMR that is dominated by electronic processes that produce broad absorption signatures. The shortwave infrared (SWIR, 1.0—2.5 μm) range is reflected EMR that is dominated by overtone and combination modes of the foundation vibration in molecules that produce sharp absorption signatures. The thermal infrared (TIR, 3.0—14.0 μm) range—that is split into the mid-wave infrared (MWIR, 3.0—5.0 μm) and the long wave infrared (LWIR, 8.0—12.0 μm) ranges—is emitted with some reflected EMR that is dominated by the fundamental molecular vibration processes, which produce both sharp and broad absorption signatures. All of these measurements use both sun (VNIR-SWIR) and earth (TIR) energy sources, characterized by Plank’s equations, and thus are termed passive.

HRS technology provides high spectral resolution and expands remote geology sensing into a new era. The precursors of HRS technology already exist in many other disciplines, and the achievements of geology in mineral recognition and mapping are vast. A very comprehensive review can be found in Schaepman et al., (2006), in van der Meer et al. (2012) and in Kruse (2012). Remarkable progress in hyperspectral technology, using both optical and thermal sensors, has been achieved since the first paper of Kruse and Kierein-Young (1990). They demonstrated a multisensory approach (AVIRIS, TIMS and SAR) to derive the surficial composition of rocks and soils. Today, as new TIR sensors are entering the market, HRS capabilities are expanding to new dimensions (Kruse 2015). By recognizing silicate-bearing minerals, which cannot readily be discerned in the optical region, Notesco et al. (2015) have demonstrated the advantages of mineral mapping using the TIR region. They mapped Makhtesh Ramon, in the Negev region of Israel, and obtained high-accuracy geological maps of the area. Taking this challenge further, Notesco et al. (2016) demonstrated that mineral mapping accuracy increases as a result of merging both optical and thermal regions together, rather than processing each of them separately. The use of the LWIR region was also found to contribute additional information about soils (Kopačková et al. 2017). HRS technology has a proven capability of distinguishing between similar rocks that are only slightly different in their chemical and physical properties. For example, carbonate minerals can easily be distinguished by the CO3 overtone in the SWIR region (around 2.33 μm). Calcite minerals are characterized with an absorption feature at 3.330—2.340 μm and dolomite is characterized at 2.312—2.322 μm (Clark et al. 1990). An advanced system to capture small spectral changes in similar scenarios has recently been developed (van der Meer et al. 2018).

The use of high spectral resolution data enables the estimation of properties by exploiting the spectral feature characteristics within their direct-active chromophore. In turn, it is also possible to estimate non-chromophore properties by their mutual relationship, and thus to leverage spectroscopy to assess materials remotely in a rapid and reliable way. This approach has been applied mainly on soil materials, which are complex solid systems, to predict many chemical and physical properties, and is termed “proximal sensing” (Angelopoulou et al. 2020). A proximal sensing model that connects the spectral responses with the chemical physical properties can be extracted by using artificial intelligence machine learning tools. To that end, many samples are needed, and the model is generated empirically to address the material properties of an unknown sample by using a single spectrum. As HRS technology provides a spectrum for each pixel in the image cube, proximal sensing models can be applied on a pixel-by-pixel basis to generate a spatial quantitative view of the property in question. Hence, every spectral model for a new and unknown property should be initiated at the laboratory level with a large set of samples to generate the model.

1.3 Hyperspectral Remote Sensing in Engineering Geology

In recent years, ground-based HRS has been introduced for the lithological characterization of rock slopes in the context of geotechnical implementations, combining it with RS techniques for identifying discontinuities and surface geometry mapping (Kurz et al. 2012; Buckley et al. 2013; Park et al. 2016; Kereszturi et al. 2018). The contribution of HRS lies in surface weathering and alteration investigation, shear zone characterization, and rock fall scar analyses (Stead et al. 2019). Recently, He and Barton (2021) applied hyperspectral imaging at unstable slopes in Ray mine, and detected high concentration of swelling clay montmorillonite underlies the unstable areas. Kereszturi et al. (2023) used VNIR-SWIR reflectance spectroscopy on laboratory-tested volcanic rock samples to predict rock mechanical properties. They suggest that spectroscopy is an appropriate new tool for assessing in-situ volcanic rock masses. The potential of TIR imaging for remotely detecting intact rock bridges prior to failure has recently been established (Guerin et al. 2019). However, the ability to estimate the intact rock strength by reflected hyperspectral signatures has not been demonstrated. Hence, in this paper, for the first time, we aim to empirically characterize the mechanical properties of carbonate rocks in the laboratory by using a spectral means that uses both the reflected and the emitted portions of EMR in the range of 0.4—12 μm.

2 Material and Methods

2.1 Rock Sampling

To establish empirical models between the spectral signature of the surface of carbonate rocks and their mechanical properties, we collected 175 samples of carbonate rock from several outcrops in Israel (Fig. 1a). The rock units were selected in a way that ensured a relatively wide range of uniaxial compressive strength (UCS) values. Rock blocks were drilled in the workshop to prepare cylindrical rock samples 54 mm in diameter and 110 mm long (Fig. 1b, c). Specimen ends were ground to a flatness of 0.02 mm to minimize end effects during pressing (Fig. 1d, e). Table 1 lists the number of rock samples for different lithologies.

Fig. 1
figure 1

Sample collection and preparation: a sampling locations in Israel, b rock blocks collected from a single field trip, c sample coring, d sample cutting, and e edge grounding

Full size image
Table 1 Rock samples according to lithology
Full size table

2.2 Spectroscopy of Rock Samples

We measured the surface reflectance spectra in the VNIR-SWIR range with the Analytical Spectral Devices (ASD®) spectroradiometer, covering 2151 spectral bands, using a high-intensity contact probe (Fig. 2a). We also measured the emissivity spectra of rock samples in the LWIR range with a Telops Hyper-Cam hyperspectral camera with 120 spectral bands (Fig. 2b). A thermal image (band 10.2324 μm), as obtained from the LWIR hyperspectral camera, is shown in Fig. 2c. The reflectance measurements were conducted using a contact probe assembly that has a contact area of 7 cm2 with the rock. The measurements were applied on the grounded surface of the rock’s core at 4 locations and calculated relative to the white halon reference of LabSphere®. The four reflectance measurements were averaged into a working spectrum.

Fig. 2
figure 2

Hyperspectral measurements of rock samples: a ASD spectroradiometer, b Telops Hyper-Cam hyperspectral camera, and c thermal image at wavelength 10.62 mm

Full size image

The emissivity of the rock's core was extracted from the HyperCam sensors as follows. The cores were arranged in a matrix of 40 cores per image that faced the optic of the camera by 80 ° and from a distance of 3 m. The size of each pixel in the image was approximately 4–5 mm. From the surface area of the cylinder face of each sample, a suitable region of interest (ROI) with a size of 4–6 pixels was selected. Each ROI contained pixels with as little variation as possible to obtain as representative an average of values as possible. Spectral reading was taken and the radiance of the ROI over the core surface and wall was extracted. The measurements acquired in each set of samples always used one sample from the last measurement to ensure that the extract emissivity is similar and no external factor encountered. The downwelling thermal radiance was calculated using a gold plate and the emissivity was then extracted following the procedure outlined in Notesco et al. (2019) (Fig. 2c).

2.3 Mechanical Properties of the Rock Samples

The bulk porosity (n) of the cylindrical samples was obtained from the expression n = [1–ρd/(Gsρw]]*100% using measured values of dry bulk density (ρd), water density (ρw), and assumed values for specific gravity (Gs). Values of Gs = 2.71 and Gs = 2.87 were selected in accordance with the calcite and dolomite contents (based on the HRS measurement), respectively. The initial volume of each sample was measured using a digital caliper with a resolution of 0.01 mm. The dry mass was measured using a digital scale with a resolution of 0.01 g. The water absorption (ω) was obtained from the expression ω = (Mwet—Mdry)/Mdry*100%, where Mwet is the wet mass of the sample after soaking in water for 48 h, and Mdry is the dry mass of the sample after heating in an oven at 104 °C for 24 h. The UCS of rock samples was determined using a large hydraulic compression frame, manufactured by ELE, with an axial compressive loading capacity of 3000 kN.

2.4 Data Processing and Analysis

Processing spectral domain data from complex systems such as soils and rocks requires a data-mining chemometrics approach (Ben-Dor and Banin 1995). The procedure searches for the interaction between EMR and active chemical groups (termed “chemical chromophores”). Additionally, scattering effects based on internal rock structure—mostly grain size distribution and porosity—are also evident in this region and affect the shape of the entire spectrum (termed "physical chromophores"). Although there is a strong relationship between the rock chromophores as observed in the spectral domain and the chemical/physical characteristics of the material, the correlation is not straightforward. Hence, the extraction of quantitative information on a given rock attribute using spectral information is not a simple task; sophisticated data-mining methods must be applied to uncover this relationship. Since the goal of this study is to derive a spectral model for practical use, it is crucial to find the optimal model for a given population, and not just a rough correlation (Carmon and Ben-Dor 2017). For this purpose, we use the PARACUDA-II® engine (Ben-Dor et al. 2015). This engine has three major modules, each with a specific purpose in the modeling process: (1) outlier detection and elimination; (2) preprocessing and transformations, based on the “all-possibilities-approach”, in which all possible spectral combinations are evaluated in the preprocessing implementation; and (3) model development and validation, based on partial least squares regression and a conditional Latin hypercube sampling algorithm (Carmon and Ben-Dor 2017). We divided the population into calibration (n = 157–161) and validation (n = 36–37) groups and extracted the feature importance that a Partial-Least-Squares Regression (PLSR) algorithm elected to define the spectral-based model. The following statistical parameters were calculated to judge the performance of each model: Coefficient of Determination (R2), Standard Error of Calibration (SEC), Standard Error of Prediction (SEP), and Ratio of Performance to Interquartile Range (RPD).

3 Results

3.1 Mechanical Properties of Rock Samples

The values of the mechanical properties obtained from laboratory measurements on the rock samples are presented in Fig. 3 for dry density, porosity, water absorption, and UCS. Most of the rock samples have dry densities around 2.6 g/cm3, which is typical for carbonate rocks. The correlation between mechanical properties is shown in Fig. 4 with lithology distinction. It is apparent that chalk lithology has low strength and relatively high water absorption, while limestone and dolomites have high strength and relatively low water absorption. Both UCS and absorption are well correlated with rock dry density.

Fig. 3
figure 3

Distribution of the tested mechanical properties: a dry density, b porosity, c water absorption, and d uniaxial compressive strength

Full size image
Fig. 4
figure 4

Correlation between mechanical properties of rock samples: a UCS vs. dry density, and b water absorption vs. dry density

Full size image

3.2 Spectral Signatures of Rock Samples

The surface reflectance spectra in the VNIR-SWIR range are shown in Fig. 5 for different lithologies: a) limestone, b) chalk, c) dolomite, d) chalky-limestone. The graphs in the left column show the relative reflectance before the continuum removal (CR) process (Clark and Roush 1984), while the graphs in the right column show the normalized spectra after the CR process. The average spectra of each lithology are plotted in bold. The emissivity spectra of rock samples in the LWIR range are shown in Fig. 6 for different lithologies: a) limestone, b) chalk, c) dolomite, d) chalky-limestone, both before (left column) and after (right column) CR. Table 2 and 3 provide the main absorption features encountered in all samples across the VNIR-SWIR and LWIR, respectively.

Fig. 5
figure 5

Spectroscopy of rock samples in the VNIR-SWIR region before (left) and after (right) continuum removal (CR), for a limestone, b chalk, c dolomite, and d chalky limestone. Bold lines represent the average spectra in each graph

Full size image
Fig. 6
figure 6

Spectroscopy of rock samples in the LWIR region before (left) and after (right) continuum removal (CR) for a limestone, b chalk, c dolomite, and d chalky limestone. Bold lines represent the average spectra in each graph

Full size image
Table 2 The main absorption features observed in the reflectance signals of all samples across the VNIR-SWIR (400–2500 nm) spectral region
Full size table
Table 3 The main absorption features observed in the emissivity signals of all samples across the LWIR (8.0–12.0 μm) spectral region
Full size table

3.3 Prediction Models

The graphs in the left column in Fig. 7 show the empirical models obtained using the PARACUDA-II® data-mining machine on the reflectance data for a) dry density, b) porosity, c) water absorption, and d) UCS. As previously mentioned, the models were trained on the calibration set consisting of 75% of the samples, and are shown here on the remaining 25% that make up the test set. The best model was obtained after performing CR preprocessing on the original reflectance spectral data before running the PLSR algorithm. The correlations between the measured and the predicted values for the tested set are shown accordingly.

Fig. 7
figure 7

Prediction models (left) and best wavelengths selected by the model (feature importance and correlograms) (right) as obtained from the PARACUDA-II® for the VNIR-SWIR region for a dry density, b porosity, c water absorption, and d UCS

Full size image

It is important to know that the spectral signatures obtained from different samples are numerically not identical to each other, even in cases where the rocks are apparently of the same composition (as shown in Figs. 5, 6). The spectral signatures identify small changes in the mineralogical composition and are also affected by the crystallographic structure of the rock and the orientation of its grains (size and shape). The analyses performed consider these changes. Therefore, in the statistical analysis different spectra will yield different predicted values of mechanical properties.

The graphs on the right-hand side in Fig. 7 show the most indicative wavelengths used to perform the prediction on the tested set (called “correlograms”) that provide the feature importance that was selected by the model. The upper banner on each correlogram shows the beta coefficient that compares the strength of the effect of each individual wavelength on the prediction model. The lower banner on each correlogram shows the individual R2 for mirroring each wavelength during the prediction process. The gray zones highlight the most dominant wavelengths in the prediction models. It can be seen that the dominant wavelengths are consistently in the range of 1750—1850 nm and in the range of 2200—2300 nm, corresponding to alumino-silicates and carbonate minerals, respectively. The prediction of the porosity is also affected by reflectance at wavelengths around 500—550 nm (Fig. 7b), indicating the presence of iron oxides. The prediction of the water absorption (Fig. 7c) is strongly aligned with the porosity features, as there is a correlation between the two properties. The results from the analyses performed on the emissivity spectra in the LWIR spectral region are presented in Fig. 8. Here, again, the graphs on the left-hand side show the prediction models performed on the test set. The graphs on the right-hand side show the correlograms. The density (Fig. 8a) is affected by Si-bearing minerals (8—9 μm) and carbonates (10—11.5 μm). The porosity (Fig. 8b) is affected by 2:1 clay mineral (9 μm) and carbonates (11 μm) as well as water molecules.

Fig. 8
figure 8

Prediction models (left) and best wavelengths selected by the model (feature importance and correlograms) (right) as obtained from the PARACUDA-II® for the LWIR region for a dry density, b porosity, c water absorption, and d UCS

Full size image

3.4 Grain Size

To examine the effect of grain size on the spectral signature, we crushed a single rock chalk sample and passed its powder through a series of sieves (Fig. 9a). We scanned the material remaining above each sieve using the ASD spectroradiometer. The reflection results are shown in Fig. 9b. In order to compare between the shapes of the spectral signatures, we performed CR process, the results of which are shown in Fig. 9c. An enlargement of the range of wavelengths between 2250 and 2400 nm that are indicative of carbonate minerals is shown in Fig. 9d.

Fig. 9
figure 9

Grain size effect on chalk sample: a crushed material on sieves, b relative reflectance of different grain sizes, c relative reflectance after CR, and d zoom-in on the 2250—2400 nm region

Full size image

The reflectance signatures of the crushed material from sieves no. 100 and 200 are the same. This means that the degree of reflection reaches its maximum at sieve no. 100 and below this grain size (0.15 mm) there is no effect on the degree of light reflection. Finally, the spectral signature of the rock surface (black line) is similar to the spectral signature of crushed material with a grain size of 0.42 mm (yellow line with rectangle symbol- sieve no. 40).

4 Discussion

4.1 The Effect of Grain Size

It can be observed from Fig. 9 that the shape of the spectral signature remains similar when scanning the same mineralogy with different grain sizes. The similarity in the shape of the spectral signature suggests that it is mainly controlled by the mineralogy, while the grain size affects the albedo; that is, the smaller the grain size, the more light it reflects, but this does not affect the location and the minima of the bell-shape curves in the spectra. The statistical models that relate the spectral signature to the mechanical properties are mainly influenced by the shape of the spectral signature. Therefore, observation of the effect of the grain size supports our assumption that it will also be possible to evaluate mechanical properties at other roughness levels of the same rock sample.

4.2 Spectroscopy of Fracture Planes vs. Grounded (Outer Surface) planes

The prediction models presented in the results section were created based on the spectroscopy of the outer grounded surfaces. Since the rock is an inhomogeneous material, the outer surface does not necessarily represent the inside of the rock. To explore the variation in the spectroscopy, we scanned the fracture planes created due to the compression tests. We collected the fractured fragments after the UCS tests and scanned the fracture planes using the ASD spectroradiometer. The results of eight different rock samples are shown in Fig. 10 for different lithologies. In each graph, the spectral reflectance of the outer grounded surface is plotted with a black line, and the spectral reflectance of the fractured planes is plotted with a dashed line. It can be observed that the shape of the signatures remains very similar, except for in the VNIR region. This region is characterized by iron oxides. We believe that the differences are an artifact due to the preparation process of the samples. The sample preparation includes sawing and grounding of rock edges, which can leave a mineralogical imprint expressed in iron oxides.

Fig. 10
figure 10

Spectral reflectance of the fracture planes (dashed lines) compared to the outer grounded surface planes (full line) for a limestone, b chalk, c, dolomite, and d chalky limestone samples

Full size image

Relative reflectance ratios, namely the reflectance of the fracture planes divided by the reflectance of the grounded outer surface, are shown in Fig. 11 (both reflectances are calculated using CR processing). Values of a ratio above 1 indicate that the reflectance of the fracture planes is greater than the reflectance from the outer surfaces, while values of a ratio under 1 indicate the opposite. It is clear that in the VNIR region the reflection from the fracture planes is greater than from the outer surface for all rock samples. As mentioned above, we suggest that this is an artifact caused by the preparation procedure. On the other hand, there is no consistency in the SWIR region: in some samples the ratio is greater than 1 and in some samples it is lower than 1, regardless of the lithology or the values of the mechanical properties. Notwithstanding, in the vast majority of the samples the ratio variations do not exceed 10%, which indicates a similarity between the signatures of the fracture planes and the outer surface.

Fig. 11
figure 11

Spectral reflectance of the fracture planes divided by spectral reflectance of the outer grounded surfaces (both after CR)

Full size image

4.3 Implementation—Adjustment to Aggregates Code

We examined whether the model is able to predict the suitability of rock samples for use as aggregates in the asphalt and concrete industries, according to the ISI-3 code: Mineral aggregates from natural sources (The Standards Institution of Israel 1998). The code defines values for water absorption in rock samples as follows: up to 3.5% the rock can be used for coarse aggregate for asphalt and up to 4.0% of the rock can be used for coarse aggregate for concrete. Table 4 shows the samples that we used for the test set (25% of the total samples). For each sample, the value of the water absorption is shown in percentages as measured in the laboratory and next to it is the predictive value produced by our model. For each value, the compliance with the code was checked. Finally, we checked whether there is a match in accordance with the code between the laboratory value and the predictive value based on the model. It can be seen that the percentage of success is very high. This means that this method has the potential to accurately evaluate the water absorption of rock samples and the ability to classify whether the material is suitable for use as aggregates in the construction industry.

Table 4 The adjustment of the water absorption to the Israeli code (The Standards Institution of Israel 1998)
Full size table

4.4 Drawbacks and Further Research

The main disadvantage of a passive RS method, such as light reflections and spectral emissivity, is that the information obtained is from the surface of the object. That is, we assume that the surface of the rocks we scan represents its interior. It is clear that it will not always be possible to conclude about the properties of the material based on surface measurements. However, we assume these methods can be relevant along fresh rock outcrops, such as active quarries, or scanning aggregates that have just been quarried.

To develop any RS method, the first step is to test it in a laboratory under controlled conditions. In our study, we tested the method using a point sensor covering an area of 7 cm2 on the rock sample. Our results, presented in this paper, confirm the method's suitability.

The next step is to test the method in field conditions using HSR cameras. These cameras capture images with lower spectral and spatial resolution. The challenge will be to identify the most important wavelengths for predicting material properties and to determine the maximum distance at which spatial information is still sufficient for identifying mechanical properties accurately.

Another challenge in the field is the presence of impurities that can mix spectral pixels with other components like soil, vegetation, and dust. Addressing these challenges requires further study. This paper lays the groundwork for future research in this direction.

5 Summary and Conclusions

Our pioneering research examined the possibility of predicting the mechanical properties of carbonate rocks by means of hyperspectral sensing. We scanned rock samples using hyperspectral sensors in the optical and thermal ranges. Then, we determined in the laboratory the mechanical properties of the rock samples: density, porosity, water absorption, and UCS. The results indicated a good correlation between the measured mechanical properties and the emissivity at wavelengths that characterize carbonate and alumino-silicates minerals. The models in both the VNIR-SWIR and LWIR regions manage to predict the density, porosity and water absorption very well. That is because the mineralogy and the internal crystallographic structure control the properties that are reflected in the spectroscopy. The predictability of the rock strength is slightly weaker because this property is also controlled by internal flaws that are not necessarily detected in spectroscopy. However, the results of the models for the rock strength are definitely satisfactory.

The next step will be to study the rock surfaces using hyperspectral imaging, first under laboratory conditions and later in the field. Based on the results of the current study, we learn that the grain size and the nature of the roughness do not have a significant effect on the shape of the spectral signature. Therefore, we assume that these characteristics will not seriously affect the ability to evaluate mechanical properties using hyperspectral sensors in the field.

The ability to assess the mechanical properties of rocks through RS can improve the fieldwork of an engineering geologist, especially in places where access to rock walls is limited or dangerous. In addition, the proposed method can improve processes related to determining the mechanical properties of quarrying material in the construction and paving industries.