Spatial Mapping of Pavement Distress Using Digital Image Data and a CNN-Assisted Clustering Framework

Research on road-condition detection and mapping can be broadly categorized into three major streams:

(A) traditional image-processing and statistical approaches,

(B) deep learning–based detection and segmentation, and

(C) representation learning and clustering approaches for infrastructure imagery.

2.1 Traditional image-processing and statistical approaches

Before the widespread adoption of deep learning, pavement inspection relied primarily on classical image-processing techniques such as thresholding, morphological filtering, edge detection, and texture analysis [16]. Although computationally efficient, these methods are highly sensitive to illumination variation, occlusion, and surface irregularities, limiting their robustness in real-world settings. To address these limitations, hybrid approaches combining multiple handcrafted descriptors — such as Local Binary Patterns and Gabor filters — were proposed to improve detection accuracy. However, their performance remained dataset-dependent and difficult to generalize across different road types and image acquisition conditions.

In parallel, clustering analysis has been widely applied to structured numerical and mixed-type infrastructure data to support decision-making processes. Partition-based and hierarchical methods — including k-means, k-medoids, and agglomerative clustering — have been used to classify road segments based on traffic intensity, maintenance cost, or environmental performance indicators [17-20]. Schubert and Rousseeuw [11] demonstrated that k-medoids performance can be substantially improved through refined medoid initialization and optimized swap strategies, reducing computational complexity while maintaining clustering accuracy. Despite their effectiveness for structured indicators, these clustering approaches are not designed to handle high-dimensional image data directly and therefore remain disconnected from image-based road damage detection pipelines.

2.2 Deep learning-based detection and segmentation

With advances in Graphics Processing Unit computing, CNNs have become the dominant paradigm for pavement defect detection and segmentation. CNN-based models have demonstrated strong performance in identifying cracks, potholes, and surface deformations under varying illumination and background conditions [6, 7, 21].

More recent studies have explored advanced architectures, including edge-aware CNNs and multi-scale feature extraction, to enhance crack segmentation accuracy [22, 23]. Transformer-based vision models have also gained attention in pavement analysis. Lu et al. [24] introduced a Swin Transformer–U-Net hybrid that significantly improved segmentation performance in cluttered and noisy environments. Comprehensive reviews highlight a growing trend toward supervised, semi-supervised, and unsupervised deep learning paradigms aimed at reducing annotation costs and improving scalability [25].

Despite these methodological advances, most deep learning–based studies focus primarily on image-level detection or segmentation accuracy. Their outputs are typically limited to pixel-level or image-level predictions and are rarely aggregated or analyzed at the road-segment or network level. Consequently, these approaches offer limited support for strategic maintenance planning and infrastructure-level decision-making.

2.3 Representation learning and clustering for spatial road-condition mapping

Representation learning–based clustering approaches aim to combine feature extraction and unsupervised grouping to uncover latent data structures. Deep embedded clustering and its variants integrate representation learning with clustering objectives to produce compact and discriminative feature embeddings [26]. Extensions incorporating autoencoder architectures have demonstrated improved stability and clustering performance, particularly for high-dimensional data [27].

In industrial inspection and civil infrastructure contexts, unsupervised and weakly supervised representation learning has been shown to reduce dependence on labeled data while improving scalability and operational feasibility [28]. However, applications of representation learning and clustering to ground-level road imagery remain limited. Existing studies predominantly focus on satellite imagery or non-road infrastructure domains, such as land-cover classification and bridge damage detection [29].

Moreover, few studies explicitly link image-based damage detection with segment-level aggregation and network-scale road-condition mapping. This gap highlights the need for integrative frameworks that connect CNN-based image analysis with interpretable clustering and spatial summarization mechanisms. The ICCF framework proposed in this study addresses this gap by emphasizing the operational integration of image-level detection, structured aggregation, and clustering-based analysis for road-condition mapping.

The ICCF scheme was implemented and evaluated using real-world road condition image data collected from 319 road segments in Sleman Regency, Yogyakarta. Each road segment contains multiple digital images captured under various lighting conditions and viewing angles, reflecting realistic operational constraints.

The three types of road damage image data used are as shown in Figure 2, namely (a) potholes, (b) corrugations, and (c) alligator cracks. The dataset utilized comprises 1335 images, which are categorized into three distinct classes.

(a) Pothole

(b) Corrugation

(c) Alligator crack

Figure 2. Example of road damage image

4.1 Convolutional Neural Network performance within the Integrative Convolutional Neural Network–Assisted Clustering Framework pipeline

In this study, the role of CNN is limited to robust feature extraction and damage categorization rather than architectural innovation. The CNN process for detecting and classifying road damage images is outlined in Figure 3.

Figure 3. The outline of the Convolutional Neural Network (CNN) process

The preprocessing stage of ICCF is a critical component for ensuring data consistency, quality, and computational efficiency prior to model training. Data augmentation is applied during training to enable the model to recognize and adapt to diverse image conditions, including suboptimal lighting and noisy inputs. This process generated approximately 9,345 image variations for the training set. Augmentation was applied exclusively to the training data; the validation and test sets were left unmodified to ensure an objective evaluation of model performance.

In the input layer, the data is interpreted as a 3-dimensional matrix (X), with the first and second dimensions representing the height (H) and width (W) of the image, while the third dimension represents the number of colors (C) in RGB or Grayscale format, as in formula (1):

$X \in \mathbb{R}^{H \times W \times C}$             (1)

Taking into account the variations in the five architectures used in this study, the input images were standardized using dimensions of 224 × 224 × 3. This size was chosen because it is lighter, allows for faster training, and has a relatively low risk of image deformation in the case of road textures. To control stability during the training process, the input data is normalized using Min-Max as in Eq. (2):

$X^{\prime}=\frac{X-X_{\min }}{X_{\max }-X_{\min }}$             (2)

The convolution layer in CNN functions to recognize patterns in images by detecting basic features such as edges, corners, and textures using filters/kernels. If there is an input like formula (1) and a kernel with the number of filters l like formula (3), then the convolution operation is like Eq. (4) with b as the bias and c as the input channel:

$F \in \mathbb{R}^{l \times l \times C}$            (3)

$Y(i, j)=\sum_{m=1}^l \sum_{n=1}^l \sum_{c=1}^C X_{i+m-1, j+n-1, c} . F_{m, n, c}+b$            (4)

where,

$Y(i, j)$: feature map value at position $(i, j)$

$X_{i+m-1, j+n-1, c}$: input value at position $(i+m, j+n)$

$F_{m, n, c}$: kernel value at position $(m, n)$

$(m, n)$: index in kernel

The output dimension is affected by stride $(s)$ and padding $(p)$ such as in Eqs. (5) and (6):

$H_{\text {out }}=\frac{H-l+2 p}{s}+1$            (5)

$W_{\text {out }}=\frac{W-l+2 p}{s}+1$             (6)

Activation layers are used to help the model recognize complex patterns in input data, such as images. In this study, the Rectified Linear Unit (ReLU) activation function is applied because it is fast, simple, and does not cause gradient saturation. ReLU converts all negative values to 0, but for positive x values, the value returns to its original value, with the activation function as formula (7):

$f(x)= \begin{cases}x, & x>0 \\ 0, & x \leq 0\end{cases}$              (7)

The pooling layer chosen in this study is Max pooling because it can reduce the image size and maintain the dominant image area. For a 2 × 2 patch, max pooling is determined according to Eq. (8):

$\begin{gathered}Y(i, j)=\max \{X(2 i, 2 j), X(2 i+1,2 j), X(2 i, 2 j+1), X(2 i+1,2 j+1)\}\end{gathered}$             (8)

The fully connected layer functions to calculate the final layer output in the CNN process, and changes the dimensions of the classifiable data linearly. In this case, the feature maps from the convolution and pooling processes are flattened into vectors in n-dimensional space according to formula (9):

$X \in \mathbb{R}^n$            (9)

The fully connected layer performs a linear transformation as in Eq. (10) and works like traditional classification:

$z=W x+b$            (10)

where,

z: output before activation

W: matrix weight

x: input vector

The output layer is the final layer in a CNN architecture, producing the final model predictions and converting high-level feature representations (from previous layers) into a format that can be interpreted for specific tasks. In this study, the Softmax activation function is used to calculate the probability of an event, and the target class is determined based on the class with the highest probability. The Softmax function for the probability of the i-th class is shown in Eq. (11):

$p_i=\frac{e^{z_i}}{\sum_{j=1}^g e^{z_j}}, i=j=1,2, \ldots, g$            (11)

CNN prediction is the class with the highest probability as Eq. (12):

$\hat{y}=\arg \max _i p_i$            (12)

The five CNN architectures analyzed in this research include the second version of ResNet with 50 layers (ResNet50V2), the second version of ResNet with 152 layers (ResNet152V2), the VGGNet consisting of 16 layers (VGGNet-16), InceptionV3, and Xception. The reasons for using these five architectures in this study are as follows.

VGGNet-16 is a classic deep learning architecture that illustrates the effect of increased network depth on classification performance. It consists of 16 learnable layers — 13 convolutional layers with uniform 3 × 3 kernels and 3 fully connected layers — arranged in a simple, modular structure. Despite having approximately 138 million parameters, its architectural simplicity makes VGGNet-16 a widely used baseline and benchmark in deep learning research.

ResNet50V2 is a deep CNN that addresses the vanishing gradient problem through residual learning with skip connections. It comprises 50 layers organized into four residual stages (3-4-6-3) using bottleneck blocks for computational efficiency. These residual connections enable effective training of deep networks, contributing to strong performance in image classification and transfer learning tasks.

InceptionV3 employs inception modules with parallel convolutions of varying kernel sizes to capture multi-scale spatial features. The architecture incorporates factorized convolutions, auxiliary classifiers, and label smoothing to improve training efficiency while reducing parameter count. This design achieves a favorable balance between high accuracy and computational efficiency, making InceptionV3 suitable for practical applications.

Xception extends the Inception architecture by fully adopting depthwise separable convolutions, separating spatial and channel-wise feature extraction. This design simplifies the network structure while improving parameter efficiency and model expressiveness. Xception delivers strong classification performance with reduced computational cost, making it well-suited for resource-constrained environments.

ResNet152V2 is a very deep residual network with 152 layers that introduces pre-activation and refined bottleneck designs. These improvements enhance training stability and enable the extraction of more complex hierarchical features. As a result, ResNet152V2 achieves higher accuracy and better generalization, making it effective for challenging computer vision tasks.

Furthermore, each architecture is evaluated based on several metrics: accuracy, recall, precision, and F1-score. In this study, all evaluation metrics are reported in percentage (%). These metrics are measured by comparing the prediction results with the actual labels. For this reason, a confusion matrix for $g$  classification is shown in Eq. (13):

$C M=\left[\begin{array}{cccc}n_{11} & n_{12} & \cdots & n_{1 g} \\ n_{21} & n_{21} & \cdots & n_{2 g} \\ \vdots & \vdots & \ddots & \vdots \\ n_{g 1} & n_{g 2} & \cdots & n_{g g}\end{array}\right]$             (13)

where, $n_{i j}$ denotes the number of objects that are actually class $i$, but are predicted to be class $j$.

Using the confusion matrix, the classification performance was evaluated in terms of true positive (TP), true negative (TN), false positive (FP), and false negative (FN). TP refers to correctly predicted positive samples, TN refers to correctly predicted negative samples, FP refers to negative samples incorrectly classified as positive, and FN refers to positive samples incorrectly classified as negative. Based on these outcomes, Accuracy, Precision, Recall, and F1-score were used as evaluation metrics. Accuracy is calculated as shown in Eq. (14). Precision and Recall are defined in Eqs. (15) and (16), respectively. The F1-score, representing the harmonic mean of Precision and Recall, is given in Eq. (17).

Accuracy $=\frac{\sum_{i=1}^g n_{i i}}{\sum_{i=1}^g \sum_{j=1}^g n_{i j}}$            (14)

Precision $=\frac{1}{g} \sum_{i=1}^g \frac{T P_i}{T P_i+F P_i}$             (15)

Recall $=\frac{1}{g} \sum_{i=1}^g \frac{T P_i}{T P_i+F N_i}$              (16)

F1 - Score $=\frac{1}{g} \sum_{i=1}^g 2 \frac{(P r)_i(R c)_i}{(P r)_i+(R c)_i}$             (17)

Based on all the provisions above, this study produces a comparison of the performance of five architectures in terms of accuracy, precision, recall, and F1-Score values for three types of road damage, as shown in Figure 4. Overall, all models demonstrate strong and consistent performance, with metric values predominantly exceeding 90%, indicating good generalization on the validation dataset.

Figure 4. Architecture performance for validation data

Among the evaluated architectures, ResNet152V2 achieved the highest performance across all four metrics, with values approaching 96%. This result is attributable to its deeper architecture combined with residual connections, which enable effective hierarchical feature extraction while mitigating vanishing gradient issues. The close alignment among accuracy, precision, recall, and F1-score further indicates balanced classification behavior without significant bias toward any specific class.

ResNet50V2 ranks second, showing performance only slightly lower than ResNet152V2. This suggests that residual learning remains highly effective even with a reduced depth, offering a favorable trade-off between model complexity and predictive accuracy. InceptionV3 and Xception exhibit comparable performance, with validation metrics in the mid-94% range. Their architectural designs, which emphasize multi-scale feature extraction and depthwise separable convolutions, respectively, contribute to robust classification results, although marginally lower than the deeper residual networks. In contrast, VGGNet demonstrates the lowest validation performance, with metrics around 91–92%. The absence of residual or advanced feature fusion mechanisms likely limits its ability to capture complex patterns, particularly when compared to more modern architectures.

Furthermore, Figure 5 illustrates the comparative performance of the evaluated CNN architectures on the testing dataset using accuracy, precision, recall, and F1-score. In contrast to the validation results, a general performance decline is observed across all models, which is expected when models are evaluated on completely unseen data and provides a more realistic assessment of generalization capability.

Figure 5. Architecture performance for testing data

ResNet50V2 achieves the best overall performance on the test data, with all evaluation metrics exceeding 92%. This result indicates that ResNet50V2 offers the most robust generalization, likely due to its balanced architectural depth that effectively captures discriminative features while mitigating overfitting. ResNet152V2 follows closely, maintaining strong performance slightly above 91%. Although deeper networks perform well during validation, the marginal decrease in testing data suggests that increased depth does not necessarily translate into superior generalization under limited or diverse testing samples. InceptionV3 and Xception show moderate performance, with metric values in the range of approximately 86–90%. Their consistent but lower results indicate stable learning behavior, though with reduced discriminative power compared to residual-based architectures. In contrast, VGG exhibits the lowest testing performance, with all metrics around 75–76%. The absence of residual connections and the relatively shallow feature representation likely limit its ability to generalize to unseen data.

Although several CNN architectures achieved comparable validation metrics, a closer examination reveals differences in generalization behavior across models. The identical validation accuracy observed in some cases can be attributed to the use of a balanced dataset and macro-averaged performance measures, which may mask subtle variations between classes when evaluated on validation subsets. In contrast, testing results exhibit larger variance across models, indicating differences in robustness and generalization capacity. Deeper architectures, such as ResNet152V2, demonstrate a tendency toward overfitting, as reflected by a noticeable performance drop when evaluated on unseen data. This behavior is likely due to the limited diversity of training samples relative to the model complexity, despite data augmentation strategies.

From a practical standpoint, model selection in the proposed ICCF framework prioritizes stability and generalization over marginal gains in validation accuracy. While deeper models offer higher representational capacity, lighter architectures such as ResNet50V2 provide a more favorable balance between accuracy, computational efficiency, and robustness. Consequently, ResNet50V2 is selected for integration within the ICCF framework, particularly for deployment in resource-constrained environments where computational efficiency and reliability are critical considerations.

4.2 Segment-level aggregation results

The different data, which has never been used before in the CNN process in the first stage of ICCF, was taken from 108 road segments in Sleman Regency, with a total of 2,319 images. The results of CNN detection and classification using the ResNet152V2 architecture were then aggregated for each road segment. For example, if $n_s$ denotes the total number of damages on the s-th road segment, and $f_d$ denotes the number of damages for the d-th damage type on the s-th segment, then the proportion of the d-th damage in the s-th segment is calculated according to Eq. (18):

$P_{s_d}=\frac{f_d}{n_s}$            (18)

This study only uses three types of road damage as a grouping, so that a three-dimensional plot can be presented using the $f_d$ value for each segment as in Figure 6. The aggregation results in Figure 6 will later be used as the basis for input for road segment mapping.

Figure 6. Profile of the aggregation level of road damage

4.3 Clustering outcomes and determining the number of groups

4.3.1 Justification of the selection of the partition method for Integrative Convolutional Neural Network–Assisted Clustering Framework for the spatial road-segment

The three main stages in cluster analysis are pre-processing, algorithm selection, and validation of clustering results. In this study, pre-processing was carried out by standardizing the data using Eq. (2). K-means, block-based k-medoids (Block-KM), modified fast k-medoids (MFast-KM), or other methods can basically be applied in the third stage of the ICCF scheme. This study uses a partition-based method to cluster objects (road segments) based on the aggregation of damage levels. This technique was chosen due to its efficiency and ability to work directly with high-dimensional data.

To select the most appropriate partitioning method, three candidate methods were evaluated using synthetic data with characteristics resembling the CNN aggregation results. The use of synthetic data in this context is intended solely as a supplementary robustness assessment of the clustering mechanism, not as a proxy for real-world road imagery features. The artificial data were arranged into three cluster scenarios, and each model was replicated fifty times. The method evaluation was based on the Adjusted Rand Index, Rand Index, Jaccard Coefficient, Hubert Statistic, and Folkes-Mallow values. The Manhattan distance between two objects (say A and B) that have the variable $p$ is as shown in Eq. (19):

$d_{A B}=\sum_{v=1}^p\left|x_{A v}-x_{B v}\right|$           (19)

Three groups of 100 objects each are observed in a bivariate normal distribution with cluster centers at (150, 150), (150, 250), and (250, 175), with each having the same standard deviation of twenty. A group with an elongated pattern that resembles a road section, where each section is cut into segments measuring, for example, 20 square meters, is shown in Figure 7.

Nine groups of 100 each are observed on two normally distributed variables. Three horizontal groups for three positions from the bottom, middle, and top, each has the same center for the first variable, namely $(150,200 \sqrt{2}, 300 \sqrt{2})$. While the center of the second variable has the same value for the three horizontal groups, at the bottom position is zero, the same in the middle position is $200 \sqrt{2}$, and the same in the top position is $300 \sqrt{2}$. Three vertical groups on the left and right sides are applied a standard deviation of 20, while for three vertical groups in the middle position, a standard deviation of 25 is applied.

Figure 8 shows that in general, the three clustering methods, namely k-means, Block-KM, and MFast-KM, achieve excellent performance in artificial data scenarios with a relatively small number of groups (3 and 5 groups), as indicated by Rand Index, Adjusted Rand Index, Hubert's Statistics, Jaccard Coefficient, and Fowlkes-Mallows Index values approaching 1.0. This indicates that the cluster structure in the data in the first two scenarios is relatively easy for these methods to learn.

Figure 7. Artificial data resembling road segments

(a)

(b)

(c)

Figure 8. Clustering performance

In the first scenario (3 groups, Figure 8(a)), MFast-KM demonstrated the highest scores for almost all metrics, followed by k-means and Block-KM with very small differences (≤ 0.002). Meanwhile, in the second scenario (5 groups, Figure 8(b)), overall performance improved across all methods, with MFast-KM again achieving good results. This indicates that the model can maintain stability and robustness when the number of clusters increases moderately.

However, significant performance differences began to emerge in the third scenario (9 groups, Figure 8(c)). ARI and JC values consistently decreased across all methods, indicating the greater modeling challenges of mapping increasingly complex clusters. In this scenario, Block-KM performed relatively well (ARI = 0.8225; JC = 0.7335; FM = 0.8432), followed by k-means, while MFast-KM experienced the sharpest decline (ARI = 0.6941; JC = 0.5802; FM = 0.7347).

Overall, these results confirm that all methods are highly effective for low to medium cluster sizes (3–5 clusters). Block-KM can adapt to increasingly complex cluster structures (9 clusters). MFast-KM excels on simpler data, but experiences a significant performance decline as the number of clusters increases. Based on these results, the Block-KM method was applied in this study. Spatial analysis was used to map the damage locations on each road segment, and this spatial information allows for more accurate identification of damage-prone points. Incorporating spatial constraints or graph-based clustering methods is an important direction for future research.

4.3.2 Determining the number of clusters

To determine the optimal number of clusters, this study employs the Silhouette coefficient, gap statistic, and Calinski–Harabasz index. These indices are widely used internal validation measures that evaluate clustering quality based on compactness, separation, and relative dispersion, without requiring external ground truth labels. The Silhouette coefficient assesses the consistency of data assignment within clusters, while the gap statistic compares within-cluster variation to a reference null distribution. The Calinski–Harabasz index complements these measures by quantifying the ratio between inter-cluster separation and intra-cluster cohesion, providing a balanced criterion for selecting the number of clusters.

For each data point (road segment) $i$, the Silhouette coefficient is defined as Eq. (20). The overall Silhouette score is obtained by averaging $s(i)$ over all data points.

$s(i)=\frac{b(i)-a(i)}{\max \{a(i), b(i)\}}$           (20)

where, $a(i)$ denotes the average distance between data point $i$ and all other points within the same cluster (intra-cluster distance). b(i) represents the minimum average distance between data point $i$ and all points in the nearest neighboring cluster (inter-cluster distance).

The variance ratio criterion (VRC) is constructed by Calinski-Harabasz, as in Eq. (21) [30]:

$V R C=\frac{\left(\bar{d}^2+A_k(n-k) /(k-1)\right)}{\left(\bar{d}^2-A_k\right)}$           (21)

where,

$A_k=\frac{1}{n-k} \sum_{g=1}^k\left(n_g-1\right)\left(\bar{d}^2-\bar{d}_g^2\right)$             (22)

The gap statistic evaluates the clustering structure by comparing the within-cluster dispersion $\left(W_k\right)$ of the observed data with that of a reference null distribution, such as in Eq. (23):

$\operatorname{Gap}(k)=\frac{1}{B} \sum_{b=1}^B \log \left(W_k^{*(b)}\right)-\log \left(W_k\right)$            (23)

where, $W_k$ is the within-cluster dispersion for $k$ clusters computed from the observed data. $W_k^{*(b)}$ denotes the withincluster dispersion obtained from the $b$-th reference dataset generated under a null model. $B$ is the total number of reference datasets.

All criteria produced the same conclusion, namely that the ideal group size to be formed is two groups, as shown in Figure 9. In the Silhouette Score, the highest value is seen at k = 2 with a value close to 0.46, which indicates that in this condition, the cluster structure has the most optimal level of compactness within the cluster and the level of separation between clusters compared to other numbers of clusters. When the number of clusters is increased from k = 3 to k = 8, the Silhouette value tends to decrease or fluctuate at a lower number, thus indicating a decrease in the quality of data separation into more heterogeneous groups.

Figure 9. Cluster evaluation

Meanwhile, the Calinski-Harabasz criterion also supports these results, where the highest score is at k = 2 with a value of more than 52. After the number of clusters increases, the value of this index is unable to exceed the value at k = 2, which means that the level of separation between clusters and the strength of the internal structure of the clusters remains best at the division of two groups.

The third evaluation, namely the gap statistic, also shows a consistent pattern, where the highest gap statistic value is found at k = 2, while at a larger number of clusters, there is a decrease or instability in the value of this metric, accompanied by a larger deviation. The combination of these three indices provides strong evidence that the road segment datasets in this study form two main groups with the most clearly separated structures.

From a practical and managerial perspective, classifying road segments into two condition groups provides a meaningful and operationally relevant framework for road authorities. In routine infrastructure management, decision-making is often constrained by limited budgets, time, and personnel. Road agencies therefore tend to prioritize actions based on broad condition categories rather than fine-grained technical distinctions that are difficult to operationalize.

In this context, clustering road segments into two characteristic groups aligns well with common maintenance planning practices, where roads are typically classified into (i) segments requiring immediate or prioritized intervention and (ii) segments that remain serviceable and can be managed through routine monitoring or preventive maintenance. This binary grouping supports rapid screening of the road network and enables decision-makers to allocate resources more efficiently without the need for complex multi-class assessments.

The ICCF scheme facilitates this process by transforming image-level damage detections into segment-level profiles that reflect the overall severity and distribution of surface defects. By grouping these profiles into two clusters, the resulting road-condition map provides a clear and interpretable representation of network-level conditions. Such representation is more actionable than descriptive tables alone, as it highlights spatial patterns of deterioration and supports strategic prioritization at the network scale.

Therefore, the use of two clusters is not intended to oversimplify road conditions, but rather to reflect the realities of infrastructure management workflows. The resulting classification also serves as a decision-support tool that bridges detailed image-based analysis and high-level maintenance planning, making it particularly suitable for use by local road agencies and public works departments operating in resource-constrained environments.

To emphasize the cluster structure in a two-dimensional form that is easy to present, the Principal Component Analysis (PCA) of the biplot is shown in Figure 10.

Figure 10. Biplot of Principle Component Analysis (PCA) for clustering

Figure 10 shows the results of applying the k-medoids method to group data into two clusters and reduce its dimensionality with PCA into a two-dimensional space (PC1 and PC2) that explains 82.3% of the data variation.

Visually, the two clusters (Cluster 1 and Cluster 2) appear quite separate, although there are still objects with close positions in the transition area between clusters. The large red dots with a cross (X) in each cluster are medoids, namely, reference objects that are the center of the group according to the characteristics of the original data (not the average value as in k-means). Cluster 1 dominates most of the area around PC1, negative to near zero, with a fairly dense distribution and a pattern that tends to form a sideways wave.

Meanwhile, Cluster 2 is located on the right side of the plot with a higher PC1 value, indicating that the main characteristics of objects in this group have a more positive PC1 score so that they are significantly different from the objects in Cluster 1. The distribution of points in both clusters shows a clearer group structure in the first dimension (PC1) compared to PC2, which indicates that the main character differences between objects are largely explained by the first principal component. Overall, the clustering results indicate the presence of two relatively consistent groups, visually and medoidally located at representative positions of each cluster, thus providing an interpretation that the separation of these two groups is quite good in explaining the pattern of data variability.

The three-dimensional plot (Figure 6) illustrates the results of clustering using the K-Medoids method based on the three original variables: the number of potholes, corrugation, and alligator damage types. This visualization also shows the distribution of road objects in three-dimensional space without transformation or dimensionality reduction, thus providing a more direct picture of the data patterns. The graph shows two clusters distinguished by different colors: the first cluster mostly includes road sections with low to moderate damage frequencies, especially in the corrugation and pothole dimensions, while the second cluster generally has higher corrugation values and is located in a denser and closer area, indicating a stronger homogeneity of characteristics. Large dots with red crosses mark the positions of the medoids of each cluster, namely representative objects that best describe the pattern of each group. The separation of the two clusters is quite visible in the combination of the corrugation and pothole dimensions, although the points in cluster 1 are more widely distributed and tend to reflect a greater variety of damage conditions.

Overall, this 3D plot confirms the presence of two main characteristics of road damage conditions: a group of roads that predominantly experience light–moderate damage and a group of roads that tend to be more severely corrugation-damaged, and these results support the visual interpretation that the cluster structure is naturally formed from the data.

4.4 Spatial interpretation and application for road maintenance

Based on the distribution map of road damage clusters (Figure 11), a spatial pattern is visible, indicating that sections included in Cluster 2 (marked with a thick red line) tend to be concentrated in the northern and northwest parts of Sleman Regency, namely the areas closer to Mount Merapi. This area is the main distribution route for volcanic materials, such as sand and stone, which are intensively transported by heavy trucks. The excessive load and frequent movement of these transport vehicles contribute to creating repeated stress and accelerating the degradation of the road pavement layer. Meanwhile, the road sections in Cluster 1 (thick black line) are spread relatively evenly across the central to southern regions of Sleman Regency, Yogyakarta, Indonesia, on road sections that are not the main routes for mining truck traffic, so the level of damage is lower.

Figure 11. Spatial pattern of road damage condition

This finding indicates that economic activities based on the utilization of Merapi's resources have a significant influence on the level of damage to road infrastructure, making mining logistics routes a top priority in planning for maintenance and improvement of pavement quality by the local government.

The interpretation of the resulting clusters is intended to support network-level screening and prioritization rather than to replace detailed engineering assessments. While individual metrics such as repair cost or severity indices are not explicitly computed in this study due to data availability constraints, the clustering results provide an intermediate decision-support layer that aggregates image-level damage information into interpretable road-segment profiles.

Compared with a simple summary table, the clustering-based mapping offers added value by revealing patterns of similarity and contrast among road segments that are not immediately apparent from descriptive statistics alone. Road segments within the same cluster share comparable damage composition and distribution characteristics, enabling road authorities to identify groups of segments likely to require similar levels of intervention or monitoring. The figures presented in this study collectively illustrate the stages and rationale for component selection within ICCF, as well as the end-to-end behavior of the proposed ICCF implementation, from image-level damage detection to network-level road condition mapping.

From an operational perspective, the clustered road-condition map facilitates rapid prioritization by highlighting segments that exhibit consistently higher concentrations of surface damage relative to the rest of the network. This information can be used as a preliminary filter to guide more detailed inspections, budgeting analyses, or cost-based evaluations conducted by road agencies. As such, the ICCF framework complements, rather than replaces, traditional engineering metrics by providing a scalable and interpretable means of translating image-based detections into actionable spatial insights.

The experimental results confirm that ICCF effectively integrates CNN-based image analysis with a clustering framework without altering the underlying algorithms. By aggregating predictions at the segment level, ICCF reduces image-level uncertainty and yields more stable and interpretable groupings of road conditions. This structured integration allows heterogeneous visual information to be consolidated into consistent representations that are meaningful at the road-segment scale.

These outcomes indicate that the main contribution of ICCF lies in its operational integration and application-oriented design rather than in methodological novelty. The framework is particularly suited to supporting spatial decision-making in road maintenance planning, where reliability, interpretability, and ease of implementation are more critical than introducing new learning algorithms.