Handling imbalanced data in supervised machine learning for lithological mapping using remote sensing and airborne geophysical data

1.1 Background

Lithology classification is critical for geological inquiry and mineral resource prospecting. On the other hand, the standard lithological mapping method is a knowledge-driven method that requires experts and is time-consuming. As a result, data-driven lithological mapping technologies, such as machine learning algorithms (MLAs) capable of rapidly processing enormous amounts of data, have been created [1,2,3]. The use of machine learning using remote sensing data combined with aerial geophysical data has been extensively researched for lithological mapping on a large scale, efficiently, and on time [4,5,6,7].

Prospecting efforts in machine learning have improved lithological mapping. Random forest (RF) is one of the most recent techniques utilized in this field, and it has shown promising results in properly predicting lithology from well logs [8] and seismic data [9]. Another technique that has gained prominence in machine learning is semantic analysis [10]. This method uses meaning from natural language documents using statistical and machine learning techniques. It has been used in image recognition, natural language processing, and speech recognition among other areas. Semantic analysis can be utilized in lithological mapping to investigate and extract geological aspects from language descriptions of lithology [11] and reflect the contextual usage meaning of the terms [12], thereby assisting in forecasting lithology types. The combination of RF with semantic analysis [13,14] can result in more accurate and efficient lithological mapping methods, which has significant implications for a variety of industries, such as oil and gas exploration, geothermal energy, and groundwater resource management.

The RF learning algorithm has been identified as a highly accurate and robust MLA for lithology classification using remote sensing data, with the ability to classify lithology types in each pixel of the study area and generate robust prediction uncertainties, making it a valuable tool for determining the spatial distribution of lithology [15]. Many studies have used MLAs, such as the RF algorithm, for lithology classification [6,7,16,17,18]. The RF algorithm is a supervised machine learning approach that is used to categorize lithology in each pixel in the study area. This algorithm is an ensemble learner that has been recognized as a good classifier, moderately robust to outliers and noise [8,19], simple-to-use, and a highly accurate approach for inferring the spatial distribution and generating robust prediction uncertainties [20]. For instance, Cracknell and Reading’s research [21] revealed that RF is straightforward to train, computationally efficient, highly stable in the face of changes in classification model parameter values, and as accurate as, or significantly more accurate than, the other MLAs tested. Meanwhile, Harris and Grunsky [22] took similar techniques. They explained how the RF algorithm can map geology and how it can reduce classification uncertainty on prediction maps.

MLA implementation necessitates the use of training samples (TSs) to carry out the “learning” process. TSs that do not have the same number of samples for all lithology classes (imbalanced) are more likely to generate bias in the classification process, favoring the classes that have the most significant or a majority of the samples. Furthermore, it complicates the identification of regularities, particularly the homogeneity patterns in the minority class [23]. However, the classifiers tend to have excellent accuracy for the majority class while obtaining poor results for the minority class [24]. In such instances, the minority class is frequently the most critical; hence measures to boost its recognition rates are required [25]. Generally, during the classification, a technique that maximizes accuracy is carried out in the training phase. In an imbalanced scenario, maximizing the overall accuracy may not be the most beneficial method [26] because, even if the prediction model yields a high overall precision, the minority class remains unknown [27,28]. Collecting TSs in lithological mapping in areas that are difficult to access is challenging. Because the number of samples is excessive and irregular at the absolute rarity level, this situation tainted the data. This suggests that the minority class lacks sufficient samples to learn the decision limits [29].

The imbalance ratio (IR) is a ratio that describes the condition of uneven data [30]. The IR is a comparison of data from the majority (negative samples) and minority (positive samples) such as 100:1 or 100,000:1. Although IR is used to sort diverse data sets, it does not always provide an accurate estimate of the difficulty level of the sample [24]. Instead, it provides the most likely class occurrence for each instance of the least likely class [31]. On the other hand, minority or rare classes are difficult to identify. This is due to the infrequency and casualness of the condition, and misclassifying rare classes results in high costs [27]. In this respect, it is crucial to make accurate predictions or identifications of the rarer rather than standard classes [29]. Furthermore, classical classifiers are prone to bias in evaluation and perform poorly for the minority class and the majority.

There are two strategies for overcoming data imbalance problems: (i) improvements at the data level and (ii) improvements at the algorithm level [24,30,32]. At the data level, the steps are to preprocess the data and rebalance the class distribution through resampling by applying the over- or undersampling techniques [32]. Solutions for dealing with unbalanced classes at the data level primarily focus on changing the class distribution to get a more balanced sample [33]. At the algorithm level, adjustment is carried out on the learning procedure, including ensemble learning approaches (e.g., RF) and cost-sensitive learning (CSL) techniques [24,32]. In addition, these two combination methods are called hybrid [25], for instance, RF-over/undersampling and RF-CSL.

Improvements at the data level include numerous techniques, for example, oversampling for tiny classes, undersampling for board categories, informative oversampling for small groups, informative undersampling for common classes, oversampling in small classes using synthetic data, and a combination of the aforementioned techniques [32]. The oversampling method works on the minority class by replicating the data to ensure it becomes balanced. The key advantage is that it minimizes information loss, but the disadvantage is that it just duplicates data and leads to overfitting [3].

CSL (also known as class weight tuning) is a method for modifying algorithms at the algorithm level [24,25,34]. This method is made of a specific set of algorithms that are sensitive to various costs associated with distinct properties of the classification problem. CSL intends to train classifiers to focus on classes with greater costs, which are prioritized. The cost of classes is one way that can be used. This scenario assumes that if the classifier incorrectly identifies a specific class, it will pay a substantial penalty/cost. The algorithm modification is performed by giving weights to the majority and minority classes. The difference in weight affects the classification during the training process. The adjustment tries to penalize misclassification in the minority group by increasing class weight and decreasing the value for the majority. According to Karlhede [35], when calculating the Gini Index, which determines the best feature in splitting data at the decision node and when the leaf acquires its class label, the provided weight should be changed. However, it should also be noted that specified weights have a threshold. When a minority class is given a very high weight, the algorithm is likely to be biased toward that group, lowering the model’s overall performance [36].

In lithological mapping, data imbalances are common. However, little research has looked into the effect of imbalances and procedures on classification accuracy, especially in small sample sizes [17,33]. The degree of imbalance may have an impact on the performance of predictive models trained on uneven data. For some samples, the imbalance can have a considerable impact on the classification quality [37]. This work investigates and analyzes multiple approaches to handling unbalanced data in lithological classification based on RF algorithms, using remote sensing and airborne geophysical data with limited borehole log samples in Komopa, Papua Province, Indonesia. These RF algorithms used oversampling and class weight tuning techniques to generate two hybrids. The oversampling method was chosen because it has successfully addressed the problem in lithological mapping [3]. The class weight tuning approach was used since it was relatively considerable, and many studies indicate successful outcomes when a large number of TSs are used [38,39]. Unfortunately, no research has used this method for lithological mapping using a limited sample collection at this time.

1.2 RF classifier

At the University of California, Berkeley, Breiman and Cutler created RFs [19,22]. This algorithm has evolved into a trustworthy and useful classifier for scientists in a variety of fields, including geology [1,40]. Numerous lithological mapping investigation has demonstrated that these techniques outperform other algorithms [1,22,41]. RF is an excellent starting point for undertaking multi-dimensional data classification. This technique can also generate a ranking list of the best predictors without requiring normally distributed TSs [42].

RF is a supervised MLA that incorporates several decision trees and mixes the decisions from many trees (ensemble classification method). To generate many decision trees, this approach uses bootstrap aggregation [19]. This approach, like other supervised classification algorithms, requires training data (i.e., locations of different rock types) [22]. Using replacements, this strategy selects two-thirds of the TSs at random to form an “in-bag” data set. The in-bag data classify a decision tree using the Gini index [43], which is used to identify the best parting for a particular class, while the remaining data (“out-of-bag” data) are utilized to validate the model. The main advantage of this algorithm is that it predicts classes based on the average of several decision trees, improving predictions and lowering outliers’ mistakes [41].

1.3 The objectives

Because of the restricted number of TSs, the problem of data imbalance necessitates research into the ideal number of samples required to achieve accurate classifications. Besides, a few study have described the behavior of minority data and its association with the IR [28]. As a result, the specific goal of this research is to determine: 1) which method produces the best classifications; 2) the effect of the number of TS datasets on the accuracy of the classifications; 3) the behavior of minority data in the classification process, and 4) the effect of the IR on the classifier’s ability. The quality of the classifications was determined not only by performance metrics computed from confusion matrices but also by a visual comparison of the lithological map obtained from the RF classification with the current map. In lithological mapping, performance indications alone are deemed insufficient to describe classification success. The visual analysis revealed how similar the distribution of lithological types and boundaries was to the previous map.

This study included the following experiments. It classifies both hybrid approaches with unbalanced data and varied amounts of data, ranging from 25 to 500 samples distributed at simple random. These studies identify the number of samples, the IR, and the distribution of the minority TSs that produce the best outcomes. Next, compare the classification result of these two hybrids to those of the balanced class weight technique and the balance TS. The balanced class weight approach assigns equal weight to each lithology class. The balanced TS, on the other hand, is a model with the same amount of TSs for each class, which is dispersed using stratified random sampling with 25 and 50 samples.

2.1 Study areas and data used

2.1.1 Study area

The research area includes the Komopa area in the Paniai District, Papua Province, Indonesia (the western part of New Guinea Island) (Figure 1). We chose this area because it already had a lithological map made by Mine Serve International (MSI) conventional geological mapping techniques to provide an accurate reference map. In 2000, MSI created a lithological map of this area with a scale of 1:25,000 [44]. Furthermore, MSI has performed a series of geophysical data collections and a large number of borehole logs so that this area is ideally used for lithological mapping trials with several geoscience data and TSs. According to the technical report written by Skead [45], the Komopa area is limited to the south by the Aga and Bogodide Rivers, both the southwest and southeast of the Wodege River, respectively. The Komopa area has moderate relief, with elevations ranging from 1,720 to 2,200 m. The terrain consists of 600–1,200-m-wide valleys filled with alluvial material divided by low hills. Longitudes 136°28′3.88″ E–136°33′54.13″ E and latitudes 3°44′38.86″ S–3°48′59.32″ S define the survey area. The majority of the land is densely forested, with trees reaching heights of 30 m, and thick humus covering the earth (Figure 2b). Quaternary alluvium, pseudo gossan, inferred porphyry, sedimentary rocks, and undifferentiated porphyry are the lithologies present in the area (as shown in Figure 2a). Pseudo gossan is a more detailed lithology class compared to others. Nevertheless, MSI used this class as a guide for their mineral exploration.

Figure 1

The area of interest is in the middle of Papua Province (Indonesia), presented by a red rectangle [48].

Figure 2

(a) A local lithological map of Komopa, scale 1:25,000, released by MSI in 2000, indicated the variation of rock types in different colors [44]. (b) Study Area, Komopa, Paniai District, Papua Province, presented by the color composite of Sentinel 2A RGB:11-08-02.

According to Glover Consulting’s geological survey reports for MSI [46], New Guinea Island is a well-known late Miocene–Pliocene copper–gold porphyry province-associated gold resource. Island arc volcanism with co-magmatic plutons is linked to mineralization. The island is formed from a collision between the Australian and the southward-migrating Pacific plates. A prominent central fold belt has formed and is split into three east–west trending structural zones: 1) the southern coastal strip comprising Palaeozoic and Tertiary shelf sediments; 2) the central mobile belt comprising Palaeozoic, Mesozoic, and Tertiary sediments, which are folded and faulted by high-angle reverse thrusts and strike-slip faults; and 3) the northern belt of Tertiary sediments and volcanic. The central mobile belt has been intruded by diorite to quartz-monzonite stocks, which are linked to skarn and porphyry copper–gold mineralization. Both intrusive and mineralization are considered structurally controlled [47].

2.2 Methodology

Figure 3 illustrates the research methodology for this study. It primarily comprises five parts: (1) preprocessing of remote sensing and geophysical data; (2) training and testing data preparation; (3) lithological classification utilizing the RF algorithm; (4) accuracy evaluation; and (5) visual comparative assessment.

Figure 3

Flowchart of research methodology to categorize lithology based on imbalanced or balanced data.

2.2.2 Training and testing data preparation

Table 2 demonstrates the number and distribution of training and testing samples used in the lithology classifications. All the imbalanced samples were acquired from simple random sampling. Meanwhile, the balance samples were obtained from stratified random sampling. Figure 4 depicts the spatial distribution of the training and testing samples.

Table 2

Training and testing sample size and distribution

Model	Number of TSs					Sum
	Inferred porphyry	Pseudo gossan	Quaternary alluvium	Sedimentary rocks	Undifferentiated porphyry
Stratified Random
Stratified Random 25	5	5	5	5	5	25
Stratified Random 50	10	10	10	10	10	50
Simple Random
Simple Random 25	2	2	2	5	14	25
Simple Random 50	2	2	3	12	31	50
Simple Random 100	4	2	8	22	64	100
Simple Random 200	5	2	12	41	140	200
Simple Random 300	6	3	16	67	208	300
Simple Random 400	7	4	19	89	281	400
Simple Random 500	10	6	23	107	354	500
	Number of Testing Samples					Sum
	8	9	17	124	344	502

Figure 4

The distribution of training and testing sample datasets with variations in the number and type of lithology. Figure (a)–(g) are TSs with a simple random distribution with a sample size variation of 25 to 500. Figure (h) and (i) depict stratified random distributions with 25 and 50 TSs. Figure (j) represents the distribution of testing samples, with a total of 502, where the class distribution is imbalanced.

The class weight tuning technique was applied using the GridSearchCV application from Scikit-learn [58]. The class weight tuning process was only applied to minority classes, such as inferred porphyry, pseudo gossan, and quaternary alluvium. Weight tuning is a search process that offers the highest training accuracy and F1 score values. Two stages of testing were performed. The first is to perform a series of classifications by giving values to the minority classes with fairly significant differences, for instance, 0.1, 1, 2, 5, 10, 15, 20. Each lithology class is given the same weight at this stage, while the resulting training accuracy and F1 score are examined. This aimed to obtain an interval of weight values with the highest training accuracy and F1 scores. When the best value interval is accomplished, the second stage of the search process is performed. This was aimed to acquire the best combination of minority class weights. In this stage, the classification is performed as in the first phase, applying the GridSearchCV application. In this application, the classification is performed by applying the value to the weight interval obtained in the first stage. The application of the weight value is conducted in stages with a relatively small increase in value. For instance, when the interval is between 0.1 and 2.0, the weight value added at each stage is 0.1.

In the oversampling approach, data preparation is performed by multiplying TSs (augmentation technique). This technique guarantees that the number equals to the highest value of TSs in the majority class. Finally, in the balance class weight method, all data points are allotted the same value, equal to 1 (known as a classic RF).

3.1 Class weight tuning results

As described in Section 2.2.2, class weight tuning results were acquired after conducting a series of classifications as described earlier. From the first stage, it was noted that the best weight value range was between 0.1 and 2.0. Afterward, in the second stage, the GridSearchCV application was applied, which computed the weights between 0.1 and 2.0 with an additional value of 0.1 at each stage utilizing the 5-fold cross-validation technique and 40,000 combinations. Table 4 displays the final results of class weight tuning.

Table 4 depicts the results of class weight tuning for the minority classes, while each majority class obtained a weight of 1. Among the weights of each class, it was discovered that not all classes have weight greater than 1. This condition implied that the weights of these classes did not surpass those in the majority class. Brownlee [63] reports that a class with greater importance was allotted greater weight, while a class with lower importance was given smaller. Meanwhile, Fernández et al. [24] mentioned that higher weights are assigned to instances emerging from the class with a higher value of misclassification cost. The minority class, with a weight of less than 1, attains less attention than the majority. Further information concerning the effects of these class weights can be observed in the performance metrics for the minor classes.

3.2 General accuracy evaluation

Table 5 demonstrates a summary of performance metrics for 23 models with varying numbers of TSs. The average score for each of the five lithology classes are displayed in this table.

Table 5 illustrates that the range of testing accuracy scores was between 0.69 and 0.80. The accuracy score increased with the number of TSs. The F1 score, alternatively, was lower than the testing accuracy. This condition is because of the classification of a limited number of outcrops and an imbalance of the sample data; therefore, the lithology prediction results from the RF algorithm are often incorrect [17]. Accuracy is suitably used when the data distribution is balanced, and when there are imbalanced classes, it is better to apply the F1 score because it signifies the actual classification performance [68].

The Kappa score values follow the same trend as the testing accuracy; hence, a more significant number of TSs will cause a higher Kappa score. The imbalance models that utilized the oversampling method produced the highest Kappa scores (ranging from 0.32 to 0.59). This Kappa score demonstrates that the level of agreement acquired is weak and minimal, with data reliability of 4–35%, which implies that more than 65% of the evaluated data contain errors [66]. The model that produced the highest Kappa score used 400 TSs (random_400_OV). For samples with a balanced distribution using 25 and 50 TSs, the Kappa scores obtained were relatively low (0.27 and 0.30), with the level of agreement reaching the category at the minimum level. These were lower than those of the imbalance models with the same number of TSs by applying the oversampling method. These findings imply that the oversampling method has enhanced the classifier’s performance.

Generally, the best F1 scores were yielded by models that applied the oversampling method, ranging from 0.43 to 0.56. These F1 scores were higher than the balance class weight model, which was the baseline model in this research, with an F1 score range of 0.37–0.52. The oversampling method could handle imbalanced data by increasing the F1 score by 0.04–0.06. Specifically, the best model was random_100_OV, utilizing 100 TSs. This number of TSs was optimal because further addition to 500 did not elevate the F1 score. Nonetheless, it tends to decrease them by approximately 0.03–0.14. This decrease in the F1 score is linked to the IR value.

3.3 Accuracy evaluation of minority class and IR

Table 6 depicts that, in general, the class weight tuning method did not offer good performance in determining minority classes. The precision and recall scores have a value of 0 in most models. Only one out of three minor classes were determined in each model analyzed. None was detected, even in the model with 100 (random_100_CS) TSs. The classification results indicate that this approach is unsuitable for categorizing lithology classes with samples that are not balanced. Nevertheless, it identified major classes, particularly undifferentiated porphyry, with the highest number of TSs (IR = 1:1) (Table 7) causing high precision, recall, and F1 score. Its F1 score ranges from 0.82 to 0.90. Another major class (sedimentary rocks) has fewer TSs (IR = 3:1), and the performance metrics scores acquired were lower than those of undifferentiated porphyry. Its F1 score ranges from 0.26 to 0.65. It was shown in the two major classes that the more the TSs were used, the higher the yielded performance metric scores were.

Generally, the balance class weight method yielded good results. Most models can determine the three minor classes (inferred porphyry, pseudo gossan, and quaternary alluvium), except for models using 25 (random_25_bal) and 200 (random_200_bal) TSs, which did not identify quaternary alluvium and pseudo gossan, respectively. Almost all classes have higher precision than recall; some even have a precision score of 1, depicting no false positive, with a low recall score (high false negative).

In general, the oversampling method provided the best classification performance in minority classes. Although the model with 200 (random_200_OV) and 400 (random_400_OV) TSs did not determine the pseudo gossan, this technique generally enhanced the F1 score in minority classes, and the overall F1 score increased compared to the balance class weight method as the baseline model. The quaternary alluvium classes that were unidentified by the balance class weight method using 25 (random_25_bal) TSs were determined by the oversampling method (random_25_OV). The model with 200 (random_200_OV) and 400 (random_400_OV) TSs, where the pseudo gossan (IR = 70:1) was not identified, showed that the oversampling method could not always identify minor classes with very high IR values. Other minority classes with an IR ≤ 40:1 could be identified. Quaternary alluvium was the minority class that consistently enhanced its F1 score across all models. This performance was possibly associated with the number of TSs in this class, which was more significant than the other two minor classes, with a lower IR value, implying that the classifier’s performance was linked to each class’s IR value. Zhu et al. [69] reported that datasets with the same IR can portray distinct classification performances when their dimensionalities vary, rendering IR suboptimal for reflecting the extent of imbalance for classification. The classification performance improved with more discriminatory features.

3.4 Visual comparative assessment

Figures 5 and 6 demonstrate visual comparative assessment results. The RF classification result was overlaid with lithological boundaries from the existing map as a visual assessment reference. There were three criteria used: (i) the ability to detect all lithology classes; (ii) the ability to classify in regards to the position and coverage area of each class; and (iii) the ability to detect lithological boundaries. To enhance our comprehension, we examine this visual comparison by detecting the amount of incorrectly identified pixels, as displayed in Tables 8 and 9.

Figure 5

Visualization of classifications from balanced models using (a) 25, (b) 50 TSs, and (c) the existing lithological map.

Figure 6

Visualization of classifications from imbalanced models using a different number of TSs. (a) Random_25_bal, (b) Random_25_CS, (c) Random_25_OV, (d) Random_50_bal, (e) Random_50_CS, (f) Random_50_OV, (g) Random_100_bal, (h) Random_100_CS, (i) Random_100_OV, (j) Random_200_bal, (k) Random_200_CS, (l) Random_200_OV, (m) Random_300_bal, (n) Random_300_CS, (o) Random_300_OV, (p) Random_400_bal, (q) Random_400_CS, (r) Random_400_OV, (s) Random_500_bal, (t) Random_500_CS, and (u) Random_500_OV.

Figure 5a and b demonstrates that the balance class weight method can determine all lithology types. In both models, it can be observed that the pseudo gossan and inferred porphyry classes are identified in a reasonably large area. Conversely, the area of the two classes does not match the area of the two related classes on the existing lithological map. This condition occurs in all classes. Table 7 shows the percentage error in the classification results. Moreover, this visualization indicated that the effect of false positives for pseudo gossan and inferred porphyry classes was visible, which results in low precision scores and high recall scores in these models (Table 5). Contrarily, these two performance metric values should have a balanced value. Nevertheless, because the testing points around the southern region are few and unevenly distributed, the two values differ (Figure 4j).

The visualization of the classifications demonstrated the performance metric scores in Tables 5, 6 and 9, which are elucidated in Figures 5 and 6. The class weight tuning method offered poor visualization results following the performance metric scores of the minority classes (Tables 6 and 9). This procedure clearly indicates the effect of the weight of the lithology class on the classifier’s ability to identify minority classes. The lithology class becomes challenging to detect at low weight values, e.g., inferred porphyry in the 25, 100, or 500 TS models. If the lithology class gets a higher weight, it will be easy to observe (e.g., quaternary alluvium in the 25, 50, or 500 models). Nonetheless, not all high weights can make the lithology class easy to detect; it can be observed in the inferred porphyry lithology class on the 200, 300, or 400 TS models. In these three models, no significant difference is seen between the inferred porphyry class, which has a high weight, and the pseudo gossan and quaternary alluvium lithology classes, which possess a low weight. This condition suggests that the weight of the class must be proportional to the weight of other classes so that the classification results do not bias toward the class with a high weight. Therefore, it can be assumed that the weights in the minor or major classes must satisfy specific proportions based on the population [70] or be inversely proportional to the frequency of the class [71]. Hence, this requires additional research to acquire a correlation between weights and weight comparisons between lithology classes.

The balanced class weight method provided better visualization than the class weight tuning method. The classification results of the 25 (random_25_bal) and 50 (random_50_bal) TS models indicate that the classifier can determine all lithology classes. However, the coverage of each lithology class does not correspond to the coverage of the related lithology class on the existing map. All lithology classes were sufficiently identified in the 100, 300, 400, and 500 TS models. Nonetheless, this procedure cannot ultimately determine the minor classes compared with the existing lithological map. For instance, in the model with 100 TSs, although the number of TS was only 2, it has a fairly low IR value (32:1). In the models with 300, 400, and 500 TS, this class was identified, although only a small area of it. Alternatively, the classifier can determine the pseudo gossan class because the number of TS for this class in each model was slightly higher (3, 4, and 6, respectively). In addition, they made the impurity value slightly more significant than the model with 200 TS, even though the IR values were almost similar (69:1, 70:1, 59:1, respectively). Regarding this, Zhu et al. [69] elucidated that their classification performances can vary when the dimensionalities of datasets with the same IR are different. Hence, IR is not the best way to demonstrate the imbalance for classification because it does not consider dimensionality.

The oversampling method offered better visual classifications than the other two methods evaluated following the three criteria described earlier. Hence, according to the visualization of the classifications, it can be noted that the model with TS of 25, 50, 100, 300, and 500 identified all lithology classes. In the model oversampling with 100 TS, it seemed that the position and area of the lithology class were entirely sufficient. Undifferentiated porphyry and sedimentary rock classes appear to be less misclassified in many locations. In this model, the quaternary alluvium was lacking in the central and southeast regions. Nonetheless, the misclassification of this class decreases with the increase of TS. The best classifications regarding position accuracy and lithology class area were generally acquired by the model with 500 TS, although the pseudo gossan class was inappropriately identified. The best identification for the pseudo gossan class was in the model with 100 TS.

The visual analysis demonstrated that lithological boundaries were accurately identified, beginning with a model with 100 TS. The classification error for the oversampling method decreases consistently as the number of TS increases, as demonstrated in Table 8. The increase in the classifier’s ability to determine the lithology class is the effect of using geophysical data. Integrating geophysical data and remote sensing is crucial to form reliable lithological maps when limited TS are utilized [41]. The sedimentary rocks generally depicted good remarks because this class has high resistivity and low magnetite, as opposed to the other four classes. This result was consistent with Zhu et al.’s study [69], which states that providing more discriminatory features can improve classification performance. The model with 500 TS portrays the best result in identifying lithological boundaries, although the pseudo gossan class was not properly identified.

The oversampling method outcomes depicted a relationship between the IR and the classifier’s ability to determine minority classes. When the IR value in the minority class was not low enough (the number of TS for both classes had an insignificant difference), then with the oversampling method, the classifier treated the minority class as though it was in a balanced state [63]. Conversely, when the IR was high (the number of TS for the majority and minority classes had a significant difference), the minority class was incorrectly identified.