Road Change Detection Using Maps and Aerial Imagery for Digital Map Updating

Abstract

Accurate and timely updating of road information is essential for maintaining reliable digital maps used in urban planning, navigation, and public services. In practical mapping workflows, road updates are typically performed by comparing existing digital maps with newly acquired aerial imagery rather than relying solely on multi-temporal imagery. This study proposes a deep learning-based framework for road change detection that integrates existing maps and high-resolution aerial orthophotos to support digital map updating. The proposed approach adopts a MapFormer-based architecture and adapts it for road-specific change detection by redefining the input and output structures to represent road-related classes. To address the severe class imbalance commonly observed in road change detection tasks, we introduce a dataset construction strategy that combines synthetic map modification and tile filtering to increase the proportion of informative change samples. In addition, different fine-tuning strategies are investigated to analyze their impact on model performance. Experiments were conducted using high-resolution aerial orthophotos and digital map data from urban areas in South Korea. The results show that incorporating synthetic data improves the model's ability to detect road changes, while excessive filtering of background samples may reduce overall robustness. Among the tested strategies, full fine-tuning of the model achieved the best overall performance. The proposed framework demonstrates the potential of map-image integration for supporting efficient and reliable digital map updating.

1. Introduction

Modern cities are in a constant state of flux due to factors such as urban redevelopment, infrastructure expansion, and changes in land use. These transformations directly influence urban spatial structures and functionality; consequently, accurately reflecting and managing these changes within digital maps is a core task for the national spatial data infrastructure. Because digital map data serves as the foundational resource for public decision-making, including traffic management, logistics, and disaster response, its accuracy and currency are directly linked to the reliability of public services.

Roads, in particular, form the backbone of urban transport networks. Beyond being simple linear features, they possess topological structures, including intersections and connectivity. Subtle disconnections or changes in road connectivity can directly impact the function of the entire network, necessitating rapid and precise updates to digital map data. Furthermore, as dynamic objects subject to frequent new construction, expansion, structural modification, and closure, it is difficult to predict the timing and scale of road changes. These characteristics underscore the necessity of a technical approach that allows for systematic and continuous change detection.

Currently, digital maps are maintained through both periodic and rapid update systems. According to Kim et al. (2024), the rapid update system is structured to collect change information from various administrative systems, which is then refined, verified, and incorporated into the digital map database. This system is systematically organized in its ability to detect potential changes from diverse information sources. However, the dualized workflow—which separates the change detection process from the data update procedure—often leads to redundant verification of identical change information.

Changes are typically subjected to a primary screening process followed by secondary validation through aerial photography, which frequently requires repetitive manual interpretation. This process increases turnaround time and hinders the timeliness of updates. Moreover, the heavy reliance on aerial photography creates a temporal lag—a “blackout” period—determined by acquisition schedules and project bidding cycles, during which changes cannot be reflected in real-time. Additionally, there are practical limitations to conducting field surveys with restricted personnel. These structural characteristics exacerbate the time lag between the occurrence of a change and its reflection in the digital map, thereby limiting the currency and reliability of spatial information. Therefore, to ensure the efficiency and reliability of map updates, an automated framework that can be integrated directly into digital map-updating workflows is essential.

To overcome these limitations, research on deep learning-based change detection has been actively conducted in recent years. Early approaches primarily utilized methods that performed object segmentation on images from each time point and then derived change areas through differencing the two results or applying rule-based post-processing. While these approaches offer the advantage of object-level interpretation, they possess structural limitations: errors generated during the segmentation stage accumulate directly in the final results, and the system depends heavily on manual post-processing rules.

To address this, end-to-end image-to-image models were proposed to directly infer change areas by taking dual-temporal images as simultaneous inputs. ChangeFormer (Bandara & Patel, 2022) improved change detection performance by effectively learning multi-scale global contextual information using a Transformer-based architecture. Additionally, SLAFormer (Lei et al., 2025) introduced a skeleton-based attention module to reflect the structural characteristics of linear objects like roads, thereby strengthening road connectivity and structural consistency. However, since these image-to-image approaches depend solely on the input images the mselves, they face limitations in sufficiently reflecting the continuity and connectivity of existing road structures, which are crucial for objects with distinct structural properties. Furthermore, in aerial orthophotos, road areas are often partially obscured, or their visual characteristics are altered by various environmental factors such as shadows, trees, and buildings. This leads to issues where static areas are misidentified as changes, or conversely, actual change areas remain undetected.

To compensate for these limitations, map–image-based approaches that utilize maps or prior information as inputs have been proposed. MapFormer(Bernhard et al., 2023) demonstrated that change detection performance could be improved through a structure that refines image features by using prior map information as a condition. Additionally, Sui et al.(2023) proposed VecRoadUpd, a method for updating road maps by analyzing discrepancies between existing vector networks and imagery. Their approach seeks more stable updates by ensuring the structural consistency of road networks and intersections.

These studies have shown that by utilizing existing road information alongside image data, it is possible to move beyond simple change detection to a more sophisticated interpretation of changes that considers structural consistency. Despite these advancements, existing map–image-based studies focus primarily on improving change detection performance or correcting specific structures, failing to expand into an integrated framework directly linked to the actual spatial information update process. Furthermore, leveraging change results for map updates requires both a clear classification of change types and the maintenance of structural consistency, yet systematic approaches to achieving both remain insufficient.

This study proposes an end-to-end, digital map–image fusion-based framework aimed at streamlining the digital map update process. Our goal is to design a system that utilizes existing digital map information alongside current aerial imagery to infer road changes by reflecting their structural characteristics, ensuring that the results can be directly applied to national rapid update systems.

The main contributions of this study are summarized asf ollows: First, this study presents a road change detection framework that integrates existing digital maps with high-resolution aerial orthophotos to support digital map updating. Unlike conventional change detection approaches based on multi-temporal imagery, the proposed approach addresses a practical scenario in which an existing map and newly acquired imagery are available for map updating. Second, this study investigates the impact of severe class imbalance in road change detection and proposes a dataset construction strategy that incorporates synthetic map modification and tile filtering to increase the proportion of informative change samples. Third, extensive experiments are conducted to analyze the effectiveness of different dataset configurations and fine-tuning strategies, providing practical insights into training deep learning models for map-guided road change detection.

2. Materials

2.1. Study Area

In this study, to objectively evaluate the learning efficiency and cross-regional generalization performance of the proposed road change detection model, we divided the study area into training and evaluation regions. The training area encompassed Anyang-si, Gyeonggi-do, and Jungnang-gu, Seoul. These regions were selected due to their heterogeneous urban structures, where residential and commercial areas are intermixed, providing diverse road densities and patterns suitable for the model to learn various urban environmental characteristics. This configuration aims to prevent the model from overfitting to specific spatial characteristics and to ensure the robustness of change detection.

Fig. 1. Geographical overview of the study area, displaying the training regions(Anyang-si and Jungnang-gu) and independent test areas(Gangnam-gu and Gangseo-gu) alongside their corresponding aerial orthophoto.

For the evaluation areas, Gangnam-gu and Gangseo-gu in Seoul were selected, which are spatially independent from the training areas. Gangnam-gu is a representative central business district characterized by a high-density urban structure with complex intersections and road connectivity, including environments where various types of road changes occur. On the other hand, Gangseo-gu features a spatial structure where large-scale planned development zones coexist with existing residential and industrial areas, showing an urban form composed of a complex mix of arterial and local roads. By utilizing two independent evaluation sites with such differing urban structural characteristics, this study aimed to comprehensively verify the model’s generalization performance across diverse urban environments, rather than limiting it to a single region. Table 1 summarizes the construction period and characteristics of the data for each region.

Table 1. Dataset overview for road change detection including acquisition periods and coverage areas

2.2. Data Sources

To perform road change detection, we utilized past vector-based digital maps and current raster-based aerial orthophotos. Historical road information was derived from 1:5,000 scale digital maps provided by the National Geographic Information Institute. These maps provide consistent vector data based on national base maps and serve as reference data for defining past road areas. For current surface conditions, we utilized aerial orthophotos with a spatial resolution of 12 cm/pixel. This data ensures high spatial accuracy through geometric correction and orthorectification, capturing intricate details such as road boundaries, intersections, and lane markings that are not reflected in digital maps.

The collected digital map road polygon datawere rasterized to match the 12 cm/pixel spatial resolution of the aerial orthophoto. To prevent data leakage due to spatial autocorrelation, the training and validation datasets were constructed in spatially separated block units within the same region. Considering the low ratio of change areas in the overall dataset, the blocks were designed and selected to ensure a sufficient number of tiles containing changes were included in the validation set. Data within each separated block was divided into 512 × 512 pixel tiles to meet the input requirements of the deep learning model. Furthermore, buffer zones were established between training and validation blocks to minimize data redundancy and spatial interference at the boundaries. This data partitioning strategy and specific spatial layout information are shown in Fig. 2. Finally, to strictly evaluate the model’s cross-regional generalization performance, data from Gangnam-gu and Gangseo-gu—areas completely independent from the training and validation datasets—were utilized as the test dataset.

Fig. 2. Spatial partitioning strategy of training and validation blocks with buffer zones to mitigate spatial autocorrelation and prevent data leakage.

2.3. Ground Truth

The ground truth (GT) for road segmentation and change detection was constructed through the following process. The road segmentation GT was generated based on the digital map’s road polygon data. Non-road objects, such as sidewalks, were excluded by performing a set subtraction operation using the sidewalk polygon layer included in the digital map, thereby defining only pure road objects, as shown in Fig. 3. The generated road polygons were registered with the aerial orthophoto and converted into a raster format corresponding to the 12 cm/pixel resolution to create pixel-wise binary labels. The change detection GT was generated by performing intersection and difference operations by comparing the past and current road raster data at the pixel level. Through this process, change types were defined asnew, removed, and unchanged areas, which were then converted into multi-class pixel labels of the same resolution to be used as training targets. The GT generation process is implemented as an automated pipeline based on spatial operations, enabling efficient construction of large-scale training datasets.

Fig. 3. Ground-truth generation process for road segmentation and change detection.

3. Methods

3.1. Overall Framework

This study proposes a framework that integrates input data configuration, model architecture, and training strategies, specifically tailored to the characteristics of road change detection. While the framework builds upon existing change detection models, it is uniquely redesigned in terms of input representation, output structure, and training strategies to directly infer the specific types of changes required for actual map updates. The overall framework consists of three primary components. First, the Dataset Configuration Strategy was developed to mitigate the issues caused by the extremely low proportion of change classes in the data. The original dataset (OD) refersto the initial training setup where aerial orthophotos are simply partitioned into 512 × 512 tiles without additional augmentation or filtering. Based on this, we applied a configuration strategy involving synthetic data generation and sample filtering to improve the ratio of change classes and enhance training stability. The specific procedures for this strategy are detailed in Section 3.4.

Second, the Map–Image Fusion Model infers structural road changes by combining prior knowledge of past road networks with current observational data, using historical digital maps and contemporary aerial orthophotos as inputs. In this model, the historical digital map is not merely a baseline for comparison but functions as conditioning information for interpreting current imagery. This design enables change inference that accounts for the connectivity and topological information of the road network. Third, the Fine-tuning Strategy is designed to adapt the pre-trained model to the specific domain of aerial-based road change detection. By employing three distinct strategies with varying parameter update ranges, we balanced the roles of the feature extraction and change determination stages, thereby improving the model's ability to discriminate between change classes. The primary contribution of this study lies in formulating and validating a practical map-guided road change detection framework for digital map updating, rather than introducing a fundamentally new generic change detection architecture. Specifically, this study adapts and reconfigures the input representation, output structure, and training strategies to better reflect the characteristics of road change detection in real-world mapping workflows. In addition, by expanding the output structure from binary change detection to multi-class classification, the model can directly distinguish between newly constructed roads, removed roads, and areas with no change.

3.2. Network Architecture

The network architecture proposed in this study is shown in Fig. 4. The network is designed to receive historical digital maps and current aerial orthophotos as inputs, inferring road change types by combining prior information on historical road structures with current observational data.

Fig. 4. Overview of the proposed map–image fusion-based road change detection framework illustrating inputconfiguration, feature fusion, and multi-class output structure.

This architecture is based on MapFormer, a Transformer-based framework that utilizes map information as a conditional input. MapFormer consists of a separate map encoder and image encoder to learn the relationship between maps and images, using an attention-based structure to combine the two features. Inthisstudy, while the basic structure of Map Former is maintained, the input representation and output structure have been redesigned to suit the road change detection problem.

Specifically, the map encoder, image encoder, and the attention-based feature fusion module are core components commonly used in both MapFormer and this study. This structure has the advantage of effectively reflecting spatial contexts, such as object boundaries and connectivity, by using map information as a condition to refine image features. For this reason, this study adopted MapFormer’s attention-basedfeaturemodulationstructure as is. This was judged to reflect the topological structure and boundary information of the road network more clearly than a simple feature concatenation method, proving advantageous for detecting changes while maintaining road connectivity, especially in complex urban environments.

Meanwhile, at the input stage, unlike the original MapFormer, which uses multi-class maps containing various land cover information, this study simplified the input into a binary format consisting of roads and background. This aims to remove semantic information irrelevant to road change detection and allow the model to focus on the road structure. The output structure also differs from the original MapFormer. While MapFormer utilizes a binary classification structure to distinguish whether a change occurred, this study expanded thisinto four classes: Background, New, Removed, and Unchanged. This design enables the direct classification of change types rather than just the presence of change, allowing for a more precise representation of the detailed aspects of road changes.

3.3. Fine-Tuning Strategies

To optimize pre-trained weights for the road change detection task, we designed three fine-tuning (FT)strategies with differential parameter update ranges based on structural consistency and experimental objectives. In all strategies, layers where direct tensor mapping with pre-trained weights was impossible due to changes in input channels orthe number of output classeswere reinitialized using standard initialization techniques.

The first strategy, FT1 (Global Optimization), updates all model parameters to optimize the model integrally for changes in the data distribution. This strategy reinitializes the initial transformation layers of the map encoder and the final output layer ofthe detection head, while setting all preceding layers, including the pre-trained image encoder, to a trainable state, thereby reconstructing the entire architecture from input feature representation to the final classification boundary for the target task.

The second strategy, FT2 (Feature Preservation), is based on the premise that the pre-trained image encoder already possesses sufficient expressive power to extract road and surface conditions. This strategy freezes the image encoder weights to preserve existing feature extraction performance while reinitializing and training only the final outputlayer oftheChange DetectionHead, whichinfers changes betweenthe input map and the current image, thereby concentrating computational resources on optimizing change detection.

The third strategy, FT3 (Bias Analysis), was designed to evaluate the bias phenomenon where pre-trained weights might adversely affect the current data distribution. This strategy reinitializes both the initial transformation layers of the map encoder and the final output layer of the detection head, while simultaneously reinitializing the entire Change Detection Head to completely exclude the semantic expressions possessed by the previous head. Consequently, by learning only the connections between the feature path containing the map information and the new Change Detection Head, we independently analyzed the impact of existing weights on task adaptability.

3.4. Synthetic Data Augmentation

In the road change detection learning process, the ratio of New and Removed classes is extremely low, causing severe class imbalance. This imbalance leads to the model becoming biased toward the Background and Unchanged classes, which results in decreased precision, especially in change boundary areas. To alleviate this, we applied a rule-based synthetic data generation technique based on the 512 × 512 tiled dataset. This technique is based on the fact that, in our model structure, the past map is utilized as reference information for comparison with the current image. As the model uses the past map as a reference representation for change inference without independently judging the authenticity of the input past map itself, artificially modifying the original past map results in inputs that function as valid training cases. The overall process of this synthetic data augmentation is illustrated in Fig. 5.

Fig. 5. Rule-based synthetic data augmentation strategy for rare change classes.

The synthetic data generation technique is composed of a rulebased method designed to maintain the topological connectivity and spatial context of real road networks. Specifically, for each 512 × 512 tile of the past map, we logically divided it into four 256 × 256 regions. To simulate New construction, we deleted the road information in three randomly selected regions, creating patterns where roads did not exist in the past but exist in the current image. Conversely, to simulate Removal, we identified background tiles where no roads existed in both the current road GT and the original past map, and artificially inserted linear structures into the past map to create patterns defined as roads that existed in the past but were removed in the present. To ensure the physical realism of the synthetic road data, we applied random variables to the length, thickness, and orientation of the added or deleted linear objects, and mitigated artificial boundary distortion through morphological operations and boundary smoothing.

Additionally, a qualitative comparison between the synthetic data and actual change cases was performed to verify whether the generated synthetic data adequately reflects real-world road change patterns; the results are presented in Fig. 6. As shown in Fig. 6, the synthetic data successfully reproduces the key structural characteristics observed in actual road changes, such as New and Removed roads. These results suggest that the proposed rule-based synthetic data generation technique effectively approximates real-world road change patterns and provides a meaningful data distribution for training the change detection model. Ultimately, this strategy increases the ratio of rare change classes while maintaining the spatio-temporal context of the original data, aiming to stabilize change boundary learning and alleviate training bias caused by class imbalance.

Fig. 6. Visual comparison of synthesized and actual road changes across multiple scenarios. The figure displays three sample cases (a–c) for each category of synthetic and real-world changes. The color-coded masks—Red (New), Blue (Removed), Green (Unchanged), and Black (Background)—highlight the spatial consistency between the rule-based synthetic data and real-world road network evolution patterns.

3.5. Loss Function

The road change detection task suffers from a significant data imbalance, as change pixels constitute only a small fraction of the total pixels, while background and unchanged pixels dominate. When learning under such rare change conditions using a conventional single classification loss, the model tends to be biased toward the majority classes (background and unchanged pixels). This imbalance leads to a disparity between precision and recall for the change class, acting as a critical bottleneck that degrades the actual road change detection performance. To mitigate this bias and perform precise road change detection, we propose a multi-loss function that combines road segmentation loss(L_sem), change segmentation loss(L_bc),separation loss(L_sep), and contrastive loss (L_con). The total loss function (L_total) is defined as Eq. (1):

L_total = L_sem + L_bc + λ_sepL_sep + L_con        (1)

Here λ_sep denotes the weight for the separation loss. In this study, the value of λ_sep was determined through experiments comparing performance variations across different weight settings, and was ultimately set to 0.2.A detailed analysis of how this weight affects model performance is provided in Section 4.2.3. The road segmentation loss(L_sem),which distinguishes between roads and the background, is defined as pixel-wise cross-entropy in Eq.(2):

\(\begin{align}L_{\text {sem }}=-\frac{1}{N} \sum_{i=1}^{N} \sum_{c \in\{0,1\}} 1\left(y_{i}^{\text {sem }}=c\right) \log p_{i}^{\text {sem }}(c)\end{align}\)       (2)

In this equation, y^sem_i denotes the ground truth road class for pixel p^sem_i represents the predicted probability for class c, and N is the number of valid pixels. Furthermore, to learn the change head, we adopted aweighted cross-entropy for the change segmentation loss (L_bc) to reflect class importance. To enhance the learning contribution of the change classes, the class weights w^bc_i were set to [1, 4, 4, 1] as shown in Eq.(3). This is intended to assign greater learning weights to the New and Removed classes, accounting for the data distribution where the proportion of change pixelsis relatively low. Given the highly imbalanced distribution where change pixels represent only a tiny fraction of the total data, this approach mitigatesthe risk of the model becoming biased toward the Background and Unchanged classes, which would occur if equal weights were used. Specifically, since the New and Removed classes appear relatively sparsely, assigning them higher weights enhances the model's sensitivity to change areas and ensures a proper balance between precision and recall.

\(\begin{align}L_{b c}=-\frac{1}{N} \sum_{i=1}^{N} w_{i}^{b c} \log p_{i}^{b c}\left(y_{i}^{b c}\right)\end{align}\)       (3)

The separation loss(L_sep)independently optimizes the change pixel set (M_chg) and the no-change pixel set (M_nochg) to prevent bias caused by the majority classes, as defined in Eq. (4):

L_sep = L_bc(M_chg) + L_bc(M_nochg)       (4)

Finally, the contrastive loss(L_con)strengthensdiscriminativepower by maximizing the separation between change and no-change embeddings within the feature space, as defined in Eq. (5):

L_con = αL^pos_t1 + βL^pos_t2 + γL^neg_t2       (5)

In this context, L^pos_t1 and L^pos_t2 denote the losses for positive pairs in the feature spaces of the reference map and the current imagery, respectively, while L^neg_t2 represents the loss for negative pairs in the current feature space. During the learning process, positive pairs consist of samples sharing the same change status, and the model is guided to minimize the embedding distance between them. Conversely, negative pairs are composed of samples with contrasting change states; by enlarging the embedding distance between these pairs, the model maximizes the characteristic discriminative power between change and no-change regions. Here, α, β, and γ are hyperparameter weight coefficients that regulate the relative contribution of each loss term.

4. Results

4.1. Experimental Setup

Model training and inference were implemented using the PyTorch framework, with the MMCV and MMSegmentation libraries utilized to establish an efficient experimental pipeline. All experiments were conducted on an NVIDIA GeForce RTX 3090 GPU. The hyperparameters for model optimization are detailed in Table 2. We employed the Adam W optimizer for weight updates, setting the learning rate to 0.00006 to ensure stable convergence. The total number of training iterations was set to 32,000, and the batch size was configured to 6, considering GPU memory efficiency.

Table 2. Hyperparameter settings used for model training

The present work designed three types of augmented datasets (AD1–AD3) by step-wise application of data augmentation and selection strategies based on the OD to alleviate the class imbalance problem in the road change detection task and improve the model’s detection performance. The OD, as the initial dataset, consists of 15,934 tiles and suffers from a severe class imbalance with an extremely low proportion of change classes, as it has not undergone any augmentation or selection processes. To resolve this imbalance, geometric augmentation, including 90-degree rotations and horizontal/vertical flips, was applied to the 201 tiles containing changes within the OD, generating a total of 905 change tiles. These tiles serve as the common change information utilized across all subsequent AD configurations.

Specifically, AD1 is an expanded configuration totaling 17,934 tiles, created by merging the OD with 2,000 tiles of rule-based synthetic data, and it includes both the previously generated augmented data and the synthetic data. Subsequently, AD2 was reconstructed into 13,486 tiles by filtering out 4,448 background tiles that contained no change information from the AD1 configuration, thereby enhancing the qualitative efficiency of the training data. Finally, AD3 is a “Change-only” configuration of 3,106 tiles that selects only those tiles containing change areasto maximize the proportion of the change class within the training data, a strategy aimed at maximizing change detection performance. Through these four designed dataset configurations (OD, AD1, AD2, and AD3), this research provides an in-depth analysis of the model’straining efficiency and the variations in road change detection performance according to each strategy. The quantitative changes in data and the pixel-level class distribution for each strategy are presented in Tables 3 and 4.

Table 3. Pixel-level class distribution across dataset configurations (OD,AD1–AD3)

Table 4. Tile-level statistics of change-containing tiles and change tile ratio across dataset configurations

Furthermore, to clarify the scope of synthetic data usage and ensure experimental fairness, the synthetic data was utilized exclusively during the training phase. The test was conducted using only real-world data, with no synthetic data included.

To systematically analyze the quantitative performance of the road change detection model, we utilized Precision, Recall, F1-Score, and Intersection over Union (IoU) as primary evaluation metrics. Each metric measures the degree of alignment between pixel-wise predictions and the ground truth. Precision and Recall are defined as shown in Eq. (6):

\(\begin{align}\text{Precision} = \frac {TP} {TP+FP}, \text {Recall} = \frac {TP} {TP+FN}\end{align}\)       (6)

Furthermore, the F1-Score, representing the harmonic mean of Precision and Recall, is defined in Eq. (7), and the IoU, which assesses the accuracy of region segmentation, is defined in Eq.(8):

\(\begin{align}\text {F1-Score}= 2\times \frac{\text {Precision} \times {\text {Recall}}}{\text {Precision + Recall}}\end{align}\)       (7)

\(\begin{align}IoU = \frac {TP}{TP + FP + FN}\end{align}\)       (8)

Here, True Positive (TP)refers to the number of pixels where the change area was correctly detected; False Positive (FP) refers to pixels where unchanged areas were misidentified as changes; and False Negative (FN) represents the number of pixels where the actual change area remained undetected.

The performance evaluation framework for this study is categorized into three aspects: class-wise evaluation, binary change detection, and road segmentation evaluation. First, the class-wise evaluation measures detection performance for specific change patterns such as new(class 1) and removed (class 2). Second, the binary change detection evaluation verifies the fundamental discriminatory power of the model in determining the presence or absence of road changes. Finally, the road segmentation results allow for an assessment of the morphological consistency and completeness of the road areas inferred by the model using the digital map data. By calculating these performance metrics for each category, we aim to provide a comprehensive analysis of how effectively the model identifies rare change classes while maintaining the precision of the overall digital map structure.

4.2. Quantitative Results

Quantitative performance evaluation comprehensively measured the change class detection capability, binary change detection, and road segmentation performance using Precision, Recall, F1-Score, and IoU metrics.

4.2.1. Dataset Strategy Comparison

Tables 5 and 6 present quantitative performance variations based on the four dataset configuration schemes (OD, AD1–AD3) defined in Section 3.1. A comparison of the results from the two evaluation areas(Gangnam-gu and Gangseo-gu) reveals that the performance trends across different dataset strategies remained generally consistent.

Table 5. Quantitative performance of dataset configurations on Gangnam-gu test area

The OD showed limitations in sufficiently learning change patterns due to the extreme data imbalance, where the proportion of change classes was very low. In both regions, it exhibited conservative detection characteristics with high precision but low recall. Specifically, in Gangnam-gu, the recall for New and Removed classes was notably low at 0.289 and 0.190,respectively, indicating a tendency to miss change areas. While Gangseo-gu showed relatively higher F1-scores and IoU, this was primarily driven by high precision in change areas, meaning the model maintained its conservative detection behavior. This behavior suggests that the relatively higher performance of OD does not necessarily indicate superior change detection capability, but rather reflects conservative predictions caused by severe class imbalance.

In contrast, AD1, which includes synthetic data, increased the proportion of change areas, allowing the model to detect changes more proactively. In Gangnam-gu, the recall for New and Removed classes improved to 0.725 and 0.519, respectively. A similar trend was observed in Gangseo-gu, where recall increased to 0.758 and 0.763. However, both regions saw a decrease in precision, which can be interpreted as an over-detection effect caused by the increased weight of change areas.

On the other hand, AD2, which applied filtering, did not lead to consistent performance improvements compared to AD1 despite the increased proportion of change pixels. In both regions, the balance between precision and recall remained limited. This suggests that, as background information was partially removed, the model failed to fully incorporate diverse spatial contexts, leading to biased learning focused strictly on change classes.

Finally, while AD3 showed high recall in some classes, it exhibited a distinct over-detection tendency with a significant drop in precision—a trend confirmed in both evaluation areas. This is attributed to the model’s inability to secure sufficient contextual information to distinguish between change and non-change in an environment where background data was excessively removed.

In conclusion, the results confirm that dataset configuration strategies influence the trade-off between precision and recall rather than providing a simple performance boost. These detection characteristics remain consistent across different urban environments. This suggests that a balanced data composition between change areas and background information is a critical factor for ensuring stable road change detection performance.

Table 6. Quantitative performance of dataset configurations on Gangseo-gu test area

4.2.2. Fine-Tuning Strategy Comparison

Based on AD1, which demonstrated the best performance among the dataset configuration strategies, we analyzed the performance variations according to the three redesigned fine-tuning strategies (FT1, FT2, and FT3).

As shown in Tables 7 and 8, FT1, which performs integrated optimization, exhibited the most stable overall performance across both evaluation regions. In the Gangnam-gu results(Table 7), it recorded a recall of 0.725 and an F1-score of 0.457 for the New class. Similarly, in the Gangseo-gu results (Table 8), it maintained stable performance with a recall of 0.758 and an F1-score of 0.419. Notably, in Gangseo-gu, the recall for the Removed class reached 0.763, confirming that change detection performance remains consistent even across varying urban structures. This suggests that organically training both the image encoder and the change detection head to suit the target data is an effective approach for road change detection tasks in diverse environments.

Table 7. Quantitative performance of fine-tuning strategies on Gangnam-gu test area

Table 8. Quantitative performance of fine-tuning strategies on Gangseo-gu test area

In contrast, FT2,the knowledge-preservation approach, showed relatively limited performance, with lower overall precision and F1-scores in both regions. This can be interpreted as a result of the fixed encoder’s feature maps being dependent on the pretraining data distribution, which restricted the information transmitted to the downstream change detection head.

Finally, FT3, the independent analysis approach, achieved high recall in some instances, but its overall F1-score remained limited due to low precision. Particularly in Gangseo-gu, while the recall for the New class was high at 0.831, the precision was low at 0.257, indicating a clear tendency for over-detection. This is likely because the pre-trained spatial context was not fully utilized due to the re-initialization of the change detection head.

Insummary, FT1, based on integrated optimization, demonstrated the most stable performance across both evaluation areas, proving its ability to maintain consistent detection performance despite differing urban structures. These results suggest that a strategy of retraining the entire model based on pre-trained representations is highly effective for stable road change detection across diverse urban environments.

4.2.3. Effect of Separation Loss Weight

To analyze the impact of the separation loss weight λ_sep on road change detection performance, experiments were conducted for λ_sep∈{0.0,0.2,0.5}, and the results are presented in Table 9. When λ_sep = 0.0, the IoU and F1-score forthe New class were 0.267 and 0.422, respectively, while the Removed class recorded 0.336 and 0.503, confirming that the overall separation performance for change areas was limited. Furthermore, the precision and recall were 0.305 and 0.685 for the New class, and 0.469 and 0.541 for the Removed class; these results indicate a tendency to detect change areas broadly with relatively high recall and low precision. This is interpreted as a result of insufficient feature representation separation between change classes when separation loss is not applied.

Table 9. Sensitivity analysis of separation loss weight (λ_sep) for change detection

Setting λ_sep = 0.2 improved the IoU and F1-score to 0.296 and 0.457 for the New class, and 0.366 and 0.536 for the Removed class, with the Removed class achieving its highest performance. In this configuration, the precision and recall were 0.333 and 0.725 for the New class, and 0.554 and 0.519 for the Removed class, demonstrating the most stable balance between precision and recall.

When λ_sep was increased to 0.5, the IoU and F1-score for the New class reached their highest values at 0.322 and 0.487, but the IoU and F1-score for the Removed class showed a decreasing trend, falling to 0.347 and 0.515, respectively. Additionally, while the precision for the New class increased to 0.390, its recall decreased to 0.650; for the Removed class, the precision remained at 0.554, but the recall declined to 0.482. This is interpreted as a result of a strengthened tendency to detect change areas more conservatively.

These results demonstrate that as λ_sep increases, the feature separation between change and no-change areas is strengthened, leading to improved precision, while the ability to detect actual change areas comprehensively decreases, resulting in lower recall. Therefore, λ_sep is a critical parameter that regulates the balance between the separation intensity and the detection sensitivity of change areas. Based on the overall performance balance and stability, λ_sep = 0.2 is determined to be the most appropriate setting in this experiment.

4.2.4. Comparison with Existing Methods

Conventional change detection research has generally focused on image-based approaches using images acquired attwo different time points. In contrast, this study addresses road change detection for digital map updating, where historical digital maps and newly acquired aerial orthophotos are jointly available. Accordingly, the proposed framework adopts a map–image input configuration and directly predicts amulti-class output consisting of Background, New, Removed, and Unchanged. These differences in input setting and output definition make direct comparison with conventional image-only binary change detection methods inherently limited.

In this problem setting, prior road map information is not an auxiliary cue but a fundamental component of the task definition. Therefore, approaches that utilize both maps and images are better aligned with the target application than methods based only on imagery, because they can explicitly reflect the existing road structure used as the baseline for map updating.

Methods that combine maps and images can be broadly categorized into post-processing-based approaches and end-to-end approaches. In our qualitative experiments on representative urban scenes, the post-processing-based approach showed that errors from the road detection stage tended to propagate to the final change results, often producing structural inconsistencies such as disconnections and false detections. In contrast, the proposed end-to-end method reduced this error accumulation by directly learning the relationship between the historical map and the current image.

For this reason, rather than relying only on direct comparison with conventional change detection methods, we additionally evaluated the practical utility of the proposed framework from the perspective of road segmentation. Although this comparison does not isolate the contribution of map input in a strict ablation setting, it provides a task-relevant assessment of whether the proposed framework can accurately reconstruct the current road network for digital map updating.

The model output is a multi-class representation of road changes, in which the New and Unchanged classes correspond to the current road area. By combining these classes, the road segmentation performance was evaluated. As shown in Table 10, the proposed framework achieved an F1-score of 0.971 and an IoU of 0.943, outperforming representative image-based semantic segmentation baselines such as SegFormer, DeepLabV3+, and UNet. These results indicate that the proposed framework can provide competitive road reconstruction performance while maintaining structural consistency.

Table 10. Comparative analysis of road segmentation performance against baseline models

Although a strict ablation that removes the map input is not fully alignedwith the practical problem setting, we acknowledge that such an analysis would be useful for isolating the contribution of map information. Therefore, future work will include controlled experiments using identical backbones with and without map input to quantify this effect further.

4.2.5. Inference Efficiency

We evaluated the inference efficiency of the proposed framework on large-scale aerial orthophotos. In Gangnam-gu, the area was partitioned into 512 × 512-pixel tiles for inference, resulting in a total of 8,526 tiles. The inference process took 476.33 seconds, which corresponds to a processing speed of 17.90 tiles per second. For the Gangseo-gu region, a total of 4,896 tiles were generated, with an inference time of 260.9 seconds, representing a speed of 18.77 tiles persecond. The results across both regions demonstrate that the proposed framework can reliably and efficiently perform inference on aerial orthophotos across different spatial scales.

Fig. 7. Qualitative examples of road change detection results. (a) Aerial orthophotos at the past time (T1). (b) Road segmentation ground truth derived from the digital map at T1. (c) Aerial orthophotos at the current time (T2). (d) Ground truth change maps indicating newly constructed roads (red) and removed roads (blue). (e) Change detection results predicted bythe proposed FT1 model overlaid on the T2 imagery, where green indicates the predicted road segmentation and colored regions represent detected change areas.

4.3. Qualitative Results

Quantitative metrics provide an objective standard by quantifying the overall performance level of the model, but they are limited in providing visual information regarding how precisely the model detects complex road boundaries or subtle structural changes. Therefore, we visualized the change detection results for actual road images, focusing on the FT1 model, which showed the most superior quantitative performance, to examine the consistency of road structures and the accuracy of boundary predictions.

Qualitative analysis results demonstrate that the proposed method maintained the shape of change areas very stably, even in environments with irregular road structures such as complex urban areas or intersections. As the knowledge from the pretrained model and the distribution of the target data were organically combined, the boundaries of change areas, such as the construction or removal of roads, showed high geometric similarity to the ground truth. This visually proves that FT1 captured linear features of roads without distortion and clearly distinguished change areas from the surrounding background by comprehensively optimizing the entire model’s parameters.

On the other hand, in areas where occlusions by trees occurred or where shadows were cast, some inference boundaries were observed to be unclear, or transient noise was present. This implies that while the model effectively learned the semantic characteristics of roads, information loss could occur during the feature extraction process in sections with strong environmental noise. Nevertheless, despite these limitations, the proposed model effectively resolved the broad false-positive problems observed in previous configuration strategies and precisely restored the overall road structure, visually supporting the excellent performance indicated by the quantitative metrics.

Consequently, qualitative evaluation confirmed that our model secured the reliability of road change detection and achieved a high level of spatial consistency in terms of semantic segmentation and geometric restoration of change areas. These visual results empirically demonstrate that the proposed fine-tuning strategy provides robust generalization performance for complex real-world road data.

5. Discussion and Conclusions

The experimental results clearly demonstrate the impact of dataset configuration strategies on model performance in road change detection tasks. Due to the extreme data imbalance, where the proportion of change areas is very low, the Original Dataset showed limited recall for change classes, exhibiting conservative detection characteristics with high precision but low recall overall. In contrast, the data configuration strategy utilizing synthetic data allowed the model to learn a wider variety of change patterns, effectively improving recall for change areas. However, as the proportion of change areas increased, a decrease in precision was observed, suggesting a clear trade-off between precision and recall in the change detection process. Furthermore, it was confirmed that excessive removal of background tiles reduces the diversity of spatial context, leading to a decline in the model's robustness. This implies that for stable road change detection, sufficient background information must be considered alongside change areas.

Regarding the comparative analysis of fine-tuning strategies, the approach of updating the entire model outperformed methods that partially froze the encoder or re-initialized the change detection head. This indicates that simply maintaining pre-trained feature representations is insufficient for complex domains like road change detection; instead, the feature extraction and decision-making stages must be adjusted simultaneously. In particular, full fine-tuning contributed to improved performance by enabling the model to more effectively capture the linear structure of roads and their surrounding contexts in complex urban environments.

Compared to existing multi-temporal image-based change detection methods, the map–image fusion framework proposed in this study offers an input configuration better suited for actual digital map update environments. In practice, existing digital maps and the latest aerial orthophotos are more frequently utilized together than temporally aligned image pairs. By using existing map information as a structural guide rather than merely a baseline for comparison, the proposed method enables interpretation of change that reflects the connectivity and topological characteristics of the road network, ensuring more stable inference even in complex urban settings.

Furthermore, this study conducted tests on two regions with different spatial scales, Gangnam-gu and Gangseo-gu, and confirmed that the performance trends based on dataset configuration strategies remained consistent across both. This demonstrates the potential for the proposed method to be applied to diverse urban environments beyond specific regions, although further validation across a broader range of areas and complex environmental conditions is required. In terms of inference efficiency, the proposed framework demonstrated stable processing of large-scale aerial orthophotos through its tile-based architecture. Since orthophotos contain spatial coordinate information, inference results performed on a tile-by-tile basis can be integrated back into a single image, enabling the generation of full-scale map updates. Moreover, the independent processing of each tile allows for scalability through parallel computing, which enhances applicability to large metropolitan areas.

Additionally, the proposed method can be extended into an automated map update system. However, in practical applications, a verification process must be integrated to account for the possibility of false positives and missed detections. Therefore, the proposed framework can serve as the foundation for an automated system supporting comprehensive map updates, while simultaneously identifying candidate change areas to improve the efficiency of subsequent update tasks.

Despite these results, several limitations remain. First, the detection performance for rare change classes, such as newly constructed roads, is still lower than that of the dominant classes. This limitation primarily stems from the severe class imbalance inherent in road change detection datasets. Second, environmental factors such as trees, shadows, and occlusion by vehicles can cause uncertainty in the detected change boundaries. To address these issues, more advanced strategies are needed to handle sparse classes and enhance the robustness of feature representations.

In conclusion, this study proposed a map–image fusion-based road change detection framework to support digital map updates. The experimental results showed that dataset configuration strategies and learning methods significantly influence change detection performance, confirming the potential of the proposed approachforpracticalmapupdateoperations.Themainsignificance of this study lies in demonstrating that a map-guided, end-to-end framework can support practical road updating more effectively than image-only formulations designed for generic change detection. Future research will aim to improve detection performance for sparse change classes and expand the scope of the framework through validation across more diverse regions and environmental conditions.

Author Contributions

Conceptualization, Data curation, Formal analysis: Lee JH; Investigation: Lee JH, Lee C, Bui NA; Methodology: Lee JH, Lee C; Software: Lee JH, Lee C, Bui NA; Visualization: Lee JH, Lee IH; Validation: Lee JH, Lee C, Cheon JW; Project administration, Supervision: Lee IP; Writing–original draft: Lee JH; Writing– review & editing: Lee JH, Lee C, Lee IH, Lee IP.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

This work is supported by the Korea Agency for Infrastructure Technology Advancement(KAIA) grant R&D program of Digital Land Information Technology Development funded by the Ministry of Land, Infrastructure and Transport(MOLIT)(Grant RS-2022-00142501).

Data Availability Statement

The data and code that support the findings of this study are available from the corresponding author upon reasonable request.