Merging Various Types of Remote Sensing Data and Social Participation GIS with AI to Map the Objects Affected by Light Occlusion

Abstract

This study proposes a practical integration of an existing deep learning model (YOLOv9-E) and social participation GIS using multi-source remote sensing data to identify asbestos-containing materials located on the side of a building affected by light occlusions. These objects are often undetectable by traditional vertical or oblique photogrammetry, yet their precise localization is essential for effective removal planning. By leveraging the mobility and responsiveness of citizen investigators, we conducted fine-grained surveys in community spaces that were often inaccessible using conventional methods. The YOLOv9-E model demonstrated robustness on mobile-captured images, enriched with geolocation and orientation metadata, which improved the association between detections and specific buildings. By comparing results from Google Street View and field-based social imagery, we highlight the complementary strengths of both sources. Rather than introducing new algorithms, this study focuses on an applied integration framework to improve detection coverage, spatial precision, and participatory monitoring for environmental risk management. The dataset comprised 20,889 images, with 98% being used for training and validation and 2% being used for independent testing. The YOLOv9-E model achieved an mAP50 of 0.81 and an F1-score of 0.85 on the test set.

1. Introduction

Some important objects affected by light occlusion cannot be detected by overhead remote sensing techniques, which is a shortcoming of these techniques. Several auxiliary methods that require a vast number of resources and budgets have been applied to elaborate on the completeness of the database.

To understand the full scope of the problem that we aim to address, it is crucial to acknowledge the pervasive nature of asbestos. Asbestos fiber has been widely used in materials since the early 20th century owing to its advantageous properties, such as softness, insulation, and low cost. However, reports from the 1930s to the 1960s linked asbestos to lung cancer, mesothelioma, and asbestosis [1,2,3]. Consequently, the IARC classified asbestos as a group 1 carcinogen in the 1970s–1980s [4], leading to its gradual reduction and global ban. Since 2000, Taiwan has progressively prohibited asbestos-containing products, with minimal production being achieved by 2019. Exposure risks persist during mining, processing, usage and demolition.

This study focuses on detecting asbestos tiles on building roofs and sides. Despite these bans, many old buildings still contain significant amounts of asbestos tiles, posing environmental and health hazards upon aging and breakage. Accurate distribution mapping is crucial for achieving effective control. Remote sensing was the primary survey method. Existing studies on roof asbestos tile distribution have utilized satellite and aerial imagery using methods such as spectral analysis, machine learning (SVM and random forest), and deep learning (CNN) [5,6,7,8,9]. Some studies have also compared model performance between SVM and CNN approaches for hyperspectral image classification [10]. In 2022, Yu et al. successfully mapped approximately 230,000 asbestos tile roofs across Taiwan and Penghu using multisource imagery and machine learning [11], thereby providing vital distribution data.

However, these studies mainly addressed roof asbestos tiles [12,13,14], overlooking the side of the building. Surveying buildings presents challenges owing to irregular shapes in imagery from varying orientations and environmental obstructions, poor lighting, and occlusions from vegetation and structures, complicating remote sensing. In contrast, manual ground surveys offer higher accuracy for irregularly oriented asbestos tiles but suffer from inefficiency and inconsistent accuracy.

To overcome these limitations, this study presents an applied integration of deep-learning models, specifically YOLO, with social participation [15]. The goal is not to propose new algorithms but to demonstrate how existing AI tools can be practically deployed to support grassroots-level environmental investigations. Mobile phone images provided by social participation were used for asbestos tile recognition with a mobile app, and location data were logged for coverage verification. The use of geotagged and orientation-aware imagery significantly improved the identification of overlooked areas when overlaid with Google Street View or Taiwan’s Building Inquiring GIS. This approach enables a more comprehensive spatial analysis of asbestos tile distribution, especially in visually restricted areas, such as vertical façades.

To structure the scope of our applied integration approach and guide our evaluations, we propose the following research questions:

• Can the integration of AI-based object detection with community-contributed imagery improve the identification of asbestos-containing materials in urban environments affected by visual occlusions?
• How effectively can participatory GIS, when combined with multi-source remote sensing data, support the spatial mapping of asbestos-containing materials in built environments?
• Is YOLOv9-E suitable for real-time inference and mobile deployment using imagery captured through social participation?

These questions reflect our core hypothesis that integrating AI models, participatory GIS, and remote sensing imagery can enhance the monitoring and mapping of asbestos-related environmental risks under occlusion and visual complexity.

2. Materials and Methods

2.1. Data Collection and Processing

This study employed YOLOv9-E as a deep learning model for image recognition of asbestos tiles. To ensure the accuracy and robustness of the AI model across diverse scenarios, we utilized a multisource image acquisition strategy incorporating Google Street View imagery, mobile phone photography, and drone photography.

Google Street View images were obtained via the Static API between January 2024 and March 2025 using parameters such as pitch values of −10 and 45, a fov value of 90, and adjusted heading angles to align with building orientations. Although precise image capture dates by Google are not always available via the API, visual inspection indicates that most images were captured between 2021 and 2024, ensuring an acceptable image currency for urban structure identification.

Mobile phone and drone imagery provided supplementary data, capturing asbestos-containing building materials from various angles under different lighting and environmental conditions. This varied dataset was aimed at enhancing the generalization ability of the model for effective and accurate identification in real-world situations. Social participation played a crucial role in rapidly and efficiently collecting a substantial volume of diverse mobile phone images, thereby significantly expanding the training dataset.

For the ground truth data, the “LabelImg” software was used for manual annotation. This process involved labeling the location and category of asbestos tiles within each image, including the position, shape, and size of the material. A data review mechanism was implemented to ensure annotation accuracy and consistency. Furthermore, the ground truth data were categorized based on the building part (e.g., roof and building side) and the image source. This refined categorization facilitates more precise training [16], enabling the model to better differentiate between various types of asbestos-containing building materials.

To maximize the utility of this diverse training dataset, advanced data augmentation techniques were applied. These included random rotations, cropping, and adjustments to brightness and contrast, effectively simulating a wider range of real-world conditions and further enhancing the model’s resilience and performance in image recognition tasks. The spatial distribution of the collected and analyzed imagery was managed and visualized using a GIS platform in Taiwan, allowing for effective spatial analysis of potential asbestos tile locations.

2.2. Establishment, Training, and Validation of Deep Learning Models

2.2.1. Deep Learning Model Selection

This study employed YOLOv9-E as an AI model for image recognition. Rather than modifying the model architecture or training structure, we adopted the YOLOv9-E (enhanced) version as is because of the balance between its high detection accuracy, computational efficiency, and suitability for mobile deployment. Developed by Wang et al., YOLOv9 integrates features such as Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Networks (GELANs), which contribute to its performance [17]. These built-in features improve gradient retention and parameter efficiency, enabling the model to operate effectively in constrained or variable imaging environments typical of citizen-captured data. Therefore, our selection of YOLOv9-E was based on its readiness for real-world application scenarios rather than architectural novelty.

It is important to clarify that this research project was initiated in early 2024 and concluded in April 2025. At the time of model selection, YOLOv9-E was released in February 2024 and offered an optimal balance among performance, efficiency, and real-time applicability on mobile devices, which aligned with our deployment goals. Although more recent object detectors such as YOLOv10 and YOLOv12 have since emerged, our model choice reflects the state of the art available during the active research period. Moving forward, subsequent studies under new funding schemes have considered the adoption of YOLOv12 or other advanced architectures. Thus, our work should be viewed as a snapshot of applied research that uses the most suitable tools available at the time of execution.

The selection of YOLOv9-E is further supported by the technical characteristics of the model. Specifically, YOLOv9 demonstrates strong performance in terms of parameter efficiency and accuracy, which are essential for applications involving variable imaging conditions and limited computational resources. The following considerations are key to our model’s adoption:

• High Accuracy and Efficiency: YOLOv9 achieves competitive detection accuracy with a relatively small number of parameters and a lower computational cost.
• Effective Learning: It offers excellent learning efficiency, even without large-scale pretraining datasets.
• Robustness: YOLOv9’s architecture supports strong generalization under diverse field conditions, including varying lighting and partial occlusion.

Key Technical Innovations

• Programmable Gradient Information (PGI) mitigates information loss in deep networks using a multilevel auxiliary information mechanism with an auxiliary reversible branch, preserving gradient flow and enhancing fine-grained features.
• GELAN supports improved information integration through hierarchical aggregation, enhanced parameter utilization, and a faster inference speed.

YOLOv9-E demonstrated superior performance in terms of parameter efficiency and accuracy compared to the other models. For this study, the YOLOv9-E (enhanced) version was selected because of its high accuracy in identifying asbestos-containing building materials.

2.2.2. Evaluation of Deep Learning Model Effectiveness

This study does not implement collaborative learning techniques such as co-training or multimodal feature fusion. Instead, it emphasizes collaboration by integrating citizen-generated data with AI-assisted detection and GIS analysis for environmental monitoring.

The performance of the trained deep learning model was evaluated by comparing its predictions with ground truth labels on a test dataset. The key evaluation metrics included the following:

• Precision: TP/(TP + FP) (minimizes false positives);
• Recall: TP/(TP + FN) (minimizes false negatives);
• F1-score: 2 * (Precision * Recall)/(Precision + Recall) (balanced metric);
• Accuracy: (TP + TN)/(TP + TN + FP + FN).

The F1-score was the primary evaluation metric, balancing precision and recall for accurate asbestos tile identification. The confidence threshold was adjusted to optimize balance. Future model improvements will focus on maximizing the F1-score to enhance the overall performance in diverse datasets and environments.

2.3. Social Participation Survey of Asbestos Tile

The pros and cons of social participation GIS (public participation GIS/PPGIS) are listed as flowing and integrating community input into spatial decision-making processes, such as urban planning, environmental management, and resource allocation.

Table 1. Summary of pros and cons of social participation GIS.

Pros	Cons
Enhanced Community Engagement	Resource Intensive
Utilization of Local Knowledge	Data Quality and Reliability
Increased Transparency and Trust	Limited Representation and Digital Divide
Improved Data Collection and Visualization	Overcrowded or Unfocused Data
Tailored Participation	Skills and Expertise Required
Capacity Building and Social Learning	Potential for Oversimplification

provides a structured overview of the key advantages and limitations of social participation GIS as identified in relevant studies.

The key consideration in applying social participation GIS is that combining targeted invitations with open recruitment can improve representativeness. Using multiple engagement methods (digital and analog) helps include diverse groups and mitigates digital divide issues. Involving experts (e.g., anthropologists) in the design and interpretation of PPGIS can help address issues of bias and data validity [18]. PPGIS is a powerful tool for democratizing spatial decision-making, but its effectiveness depends on careful planning, inclusivity, and a critical analysis of the data collected. This study addressed the challenge of efficiently mapping asbestos tile distribution on the sides of a building by integrating field investigations with social participation. Although mobile phone photography offers advantages over Street View with a higher resolution and access to confined spaces, it traditionally involves high labor costs. Leveraging Taiwan’s strong history of social participation, this research engaged the public to enhance the survey scope and efficiency and to raise public awareness. This approach goes beyond traditional policy promotion, actively involving citizens in data collection and analysis to better understand asbestos tile distribution across Taiwan.

This study incorporated social participation across two key phases, integrated with a deep learning (YOLO)-driven methodology:

• Initial Data Collection: A small group of trained citizens collected images of potential asbestos tiles on building sides and recorded their locations, contributing to the dataset for deep learning model training.
• On-site Investigation: A mobile application powered by the trained YOLO model enables a broader group of citizens to identify and locate asbestos tiles on building sides.

A core component of this research was a mobile application designed to facilitate social participation in an asbestos tile survey. The app empowers citizens to

• Capture and upload images of potential asbestos-containing materials on building sides, directly contributing to the image recognition data.
• Record precise locations using the GPS of their phones, enabling an accurate spatial analysis within the GIS framework.
• Receive guidance and information on asbestos tile identification, ensuring data quality.
• Contribute data to a centralized database, supporting the comprehensive mapping and analysis of asbestos tile distribution in Taiwan.

Mobile applications significantly enhance the efficiency and scalability of asbestos tile surveys, thereby enabling a more comprehensive understanding of the problem and informing targeted mitigation strategies.

2.4. Mobile App Development

This research incorporated a mobile application to facilitate public social participation in identifying asbestos tiles on buildings. The core functions of the app were designed to enhance the efficiency and accuracy of data collection and reporting.

Key App Functions:

• Image Recognition: Users capture photos of suspected materials, and the app employs a deep learning model, specifically YOLO, for rapid image recognition to estimate the likelihood of the presence of an asbestos tile.
• Real-time Location: The app integrates GPS functionality to record the precise locations of photographed materials, enabling accurate spatial analysis within a GIS framework.
• Reporting: Users can upload photos, identification results, and location data, including details regarding the building or other locations. These data are then used to update the central database.

The app streamlines data collection, allowing for a more comprehensive mapping of asbestos tile distribution across Taiwan.

The data handling process of the application is illustrated in Figure 1. When a user submits a photo, the deep learning model processes the image to determine whether an asbestos tile is present. If it is confirmed, the application checks the database. Existing entries are updated, whereas new cases are assigned a unique identifier (MID) and added to the database. This process ensures efficient data management and updating.

As shown in Figure 2, the app guides users to capture images of suspected materials. The integrated YOLOv9-E model then performs real-time classification, with the results and GPS location data displayed immediately on the smartphone screen.

3. Results

3.1. Image Acquisition, Labeling, Classification, and Preprocessing

Image data from the following three sources were used as the basis to train the deep learning model: mobile phone photography, UAV imagery, and Google Street View imagery. The differences between urban and rural environments in Street View imagery were also considered, as illustrated in Figure 3. Simultaneously, the LabelImg software was used for ground truth data labeling.

Table 2. Summary of image data quantity and manually labeled asbestos materials from different sources.

	Collected Images	Manual Labels
Mobile Images	13,924	31,958
UAV Images	1153	1879
Google Street View	5812	9610
Total	20,889	43,447

lists the number of images collected and manual labels for each image source.

The dataset was split into training and validation sets in an 80:20 ratio. Additionally, 10% of non-asbestos background data was included in the training set to effectively reduce false positives. This approach allows the model to maintain comparable accuracy while enabling the use of a lower confidence threshold. The YOLOv9-E models were trained using PyTorch 2.0.1 on an NVIDIA RTX 4090 GPU. The models typically converged within 100 epochs, and increasing the number of epochs beyond this threshold did not lead to further improvements in accuracy. Furthermore, hyperparameter evolution was conducted over 300 iterations to optimize the model’s performance; however, the results indicate that the default hyperparameter settings remained optimal. As shown in Figure 3, noticeable differences were observed between UAV imagery and ground-based sources such as mobile phones and Google Street View photographs. UAV images captured from a top-down aerial perspective differ significantly in terms of the viewing angle and scale. Mobile phone images tend to present asbestos materials at a closer range and with greater clarity as they are typically taken manually at short distances. By contrast, Google Street View imagery often contains more distant views of asbestos-containing materials, resulting in smaller object sizes and increased visual complexity.

Representative prediction results for unseen test images are shown in Figure 4. These visualizations demonstrate the ability of the YOLOv9-E model to detect asbestos-containing materials under diverse conditions, including different lighting, camera angles, and building materials. The images confirm that the model can accurately localize asbestos tiles across various environments and perspectives, validating its generalization performance.

3.2. Comparison of AI Classification Accuracy Tests

YOLO-based deep learning models were evaluated using different photo sources to identify asbestos tiles on building sides. The models were trained under two classification schemes: single-category (general asbestos tile) and multicategory (UAV roof, Street View roof/side, and mobile phone roof/side).

The performance of the model was assessed using the F1-score, precision, and recall, focusing on the relationship between these metrics and confidence scores.

As shown in Figure 5, the single-category model achieved a stable F1-score of 0.60 within a confidence range of 0.298 to 0.7 (Figure 5), and the multi-category model reached an average F1-score of 0.63 with a confidence score of 0.215, with variations across categories (Figure 5b). “UAV roof” performed best at higher confidence (around 0.6).

Figure 6 illustrates the precision and recall curves of the multicategory model. “Mobile Phone-side” demonstrated high precision (approaching 1) at high confidence, but recall decreased with increasing confidence. “Mobile Phone-roof” showed unstable precision and a sharp decline in recall as confidence increased.

To conduct a comprehensive evaluation of the model’s performance, we further report standard object detection metrics, including the mean average precision (mAP) at IoU thresholds of 0.5 (mAP@50) and across a range of thresholds (mAP@50:95), along with class-wise precision–recall curves.

Figure 7 compares the PR curves of one- and five-class detection models. The one-class model achieved a mean mAP@50 of 0.585. In contrast, the five-class model, which distinguishes between image sources (e.g., UAV, mobile phone, and Google Street View), achieved a higher mean mAP@50 of 0.642. Notably, the UAV_roof and Street View_roof categories reached class-specific mAP values of 0.838 and 0.763, respectively, whereas categories such as Mobile_Phone_roof performed worse.

Figure 8 and Figure 9 show the training performance of the one-class and five-class models, respectively. While both models exhibited stable convergence of loss functions, the five-class model showed superior performance in terms of both mAP@50 (0.642 vs. 0.585) and mAP@50:95 (0.35 vs. 0.297). These results suggest that explicitly modeling the visual differences between data sources may enhance generalization, particularly in occlusion-heavy or visually complex environments.

While both models achieved acceptable precision and recall values, the five-class model demonstrated superior performance, particularly for UAV and Street View imagery. Nevertheless, precision–recall trade-offs remain evident in mobile phone-based façade images, emphasizing the need for more targeted training samples.

3.3. AI-Assisted Identification of Asbestos Building Materials

This section compares the effectiveness of two image sources for AI-assisted asbestos tile identification on the building side: Google Street View and social participation images. Google Street View provides publicly available imagery for automated processing, whereas social participation is carried out using a mobile application to capture and report suspected materials. This study analyzed performance in terms of precision, recall, and F1-score within a defined test area in Tainan City, Taiwan.

Table 3. Performance evaluation of AI recognition on two image sources.

Image Sources	Google Street View	Social Participation
Total number of images	4166	246
AI identified as asbestos	403	205
AI identified as non-asbestos	3763	0
True Positive (TP)	333	205
False Positive (FP)	70	0 *
False Negative (FN)	69	41
True Negative (TN)	931	0 *
Precision (%)	82.63	100 *
Recall (%)	82.64	83.33 *
F1-score (%)	82.73	90.91 *
Accuracy (%)	90.9	83.33 *

Note: Social participation data focused on images containing asbestos, resulting in zero True Negative (TN) and False Positive (FP) values. Therefore, recall and accuracy are identical and reflect the model’s ability to identify asbestos-containing targets, not overall performance.

summarizes the performance of the AI model using deep learning for each image source. The following is a brief discussion of the AI identification performance based on four indicators: precision, recall, F1-score, and accuracy.

• Precision: The social participation model achieved 100% precision, indicating no false positives owing to close-range targeted image capture. Google Street View has 82.63% precision, with more misjudgments likely owing to image quality and obstructions.
• Recall: Both methods showed similar recall (approximately 83%), indicating comparable sensitivity for detecting asbestos-containing materials.
• F1-Score: The social participation model (90.91%) outperformed Google Street View (82.73%), demonstrating a better balance between precision and recall.
• Accuracy: While Google Street View demonstrated an accuracy of 90.90%, the accuracy measure was not applicable for the social participation dataset because of the absence of “non-asbestos” samples. Therefore, comparisons between the two methods based solely on accuracy may be misleading. Instead, precision and recall provide meaningful insights into the performance differences between the two approaches.

Figure 10 illustrates the spatial analysis of AI-identified asbestos tiles from satellite imagery, Google Street View, and social participation images. A key observation is the consistency between roof asbestos tile locations identified via satellite and building-side locations identified by Street View and social participation, suggesting a correlation in material use.

Figure 11 illustrates the surveyors’ reach and demonstrates how social participation surveys enabled us to access areas inaccessible to both satellite imagery and Street View, thereby capturing hidden or unique asbestos tiles that AI could potentially miss.

3.4. Advantages and Challenges of Citizen Science Surveys

As evidenced by the findings in the previous section, the efficiency of asbestos tile identification using satellite imagery is constrained by environmental factors such as resolution and building occlusion. Although Google Street View (GSV) data offers extensive coverage and accessibility, its utility for detailed and effective asbestos tile identification is limited by the capture locations, driving routes, and image angles of Street View vehicles. Furthermore, GSV imagery lacks the precision required for the accurate spatial annotation of asbestos-containing materials.

By contrast, social participation can access areas that are inaccessible to Street View vehicles, enabling more detailed inspections. This facilitated the effective capture of concealed, small, or uniquely patterned asbestos-containing building materials. Another significant advantage of social participation surveys is the potential for a more precise geolocation of buildings containing asbestos materials through the integration of smartphone GPS and directional metadata associated with the captured images. Consequently, despite the relatively time-consuming nature of the survey process and the potentially lower volume of images compared to GSV, the detailed inspection capabilities and inherent geolocational accuracy render social participation a viable and valuable supplementary approach for asbestos tile detection and mapping.

In the field-based validation phase, a total of 12.5 person-hours were invested by trained research staff and volunteers, resulting in the identification of 246 asbestos roofing or siding structures. This equates to approximately 19.7 detections per person-hour. Although we did not quantify which of these instances were unobservable from satellite or Street View imagery, the detection rate illustrates the practical efficiency of incorporating community-based efforts into environmental surveillance.

Table 4. Comparative evaluation of GSV vs. social participation imagery.

Aspect	AI for Google Street View—Pros	AI for Google Street View—Cons	AI for Social Participation—Pros	AI for Social Participation—Cons
Time Efficiency	Fast, automatic batch processing of public imagery	May rely on outdated imagery and inconsistent update cycles	Real-time, fresh data collected directly from field	Highly time-consuming due to investigation, image capture, and interaction
Accessibility	No fieldwork needed; covers wide and hard-to-reach areas remotely	Street View coverage may be incomplete or obstructed by objects	Can access hidden or restricted areas not visible via Street View	Requires physical presence, travel, and permissions
Detection Accuracy	Multiple angles improve identification redundancy	Limited by image quality, resolution, and environmental obstructions	Close proximity allows detailed and accurate inspection	Risk of human error, inconsistency, and subjective judgment
Geo-Annotation Capability	Basic location metadata available (pano_id, coordinates)	Cannot manually correct or fine-tune coordinates	Precise manual geo-annotation possible at time of capture	Requires extra manual input and consistent training of users
Privacy and Ethics	Uses public images; generally low privacy risk	Cannot verify recent property changes or removals	User-controlled capture ensures relevance	Potential privacy invasion; must manage ethical protocols
Data Volume and Scalability	Can process vast amounts of imagery automatically	Data quality varies; includes redundant or noisy images	High-quality, focused data with user context	Limited scalability depending on manpower and community participation
Freshness of Imagery	Enables historical comparison and change detection	Images may be several years old, missing current conditions	Provides up-to-date imagery reflecting current field status	Collection speed slower; dependent on scheduling fieldwork

summarizes and compares the advantages and limitations (pros and cons) of using GSV imagery versus social participation imagery in asbestos tile surveys.

4. Discussion

This study integrated the YOLOv9-E deep learning model, public participation GIS (PPGIS), and multi-source remote sensing imagery to identify asbestos-containing materials (ACMs) affected by visual occlusion in various environmental contexts. The findings demonstrate that this integrated framework effectively addresses the three core research questions.

4.1. Effectiveness of AI and Community-Contributed Imagery in Occluded Environments

The first research question is as follows: can the integration of AI-based object detection with community-contributed imagery improve the identification of asbestos-containing materials in urban environments affected by visual occlusions?

These findings support our hypothesis. Close-range images captured by citizens provided high-resolution targeted data on occluded or difficult-to-access buildings. The results achieved a precision of 100% and an F1-score of 90.91%, significantly outperforming the Google Street View (GSV) images. This demonstrates that community-generated imagery can effectively compensate for the limitations of traditional remote sensing in visually restricted areas, thereby improving both detection coverage and accuracy.

4.2. Complementarity of Remote Sensing and Social Participatory GIS

The second question explored was as follows: can the combination of social participatory GIS and multi-source remote sensing data support more comprehensive spatial mapping of ACMs?

The answer was affirmative. As shown in Figure 10 and Figure 11, the combined use of satellite imagery, GSV, and smartphone-based image collection forms a complementary system that enables more complete mapping of ACM distribution. This integration revealed hotspot areas that were otherwise undetectable using a single data source. Spatial complementarity, along with a real-time feedback mechanism, enhanced the precision and coverage of the mapping results.

4.3. Applicability of YOLOv9-E for Mobile Real-Time Deployment

The third question asked was as follows: is YOLOv9-E suitable for real-time inference and mobile deployment using socially captured imagery?

This study confirmed the feasibility of deploying YOLOv9-E on mobile devices. The model achieved a test set mAP@50 of 0.81, demonstrating sufficient detection accuracy and computational efficiency. The real-time classification and GPS tagging functions of the smartphone app also significantly improved user engagement and data quality, making it a practical tool for grassroots-level community participation in data collection.

4.4. Trade-Offs and Limitations of Different Image Sources and Model Performance

Despite the effectiveness of the integrated approach, each image source presents a trade-off. GSV offers large-scale, automated processing capabilities but is limited by the image update frequency and viewing angles. In contrast, social participation imagery provides fresh and high-resolution data but requires significant human effort and time. As summarized in Table 4, these sources differed in terms of time efficiency, accessibility, detection accuracy, and annotation capacity. Practical applications should balance their use adaptively based on situational needs.

Regarding model performance, while YOLOv9-E showed overall stability, the classification strategies influenced the outcomes. The five-class model performed better for the UAV and GSV images, whereas the single-class model maintained its consistency. However, mobile phone images, particularly those of building façades, exhibited greater variability in precision and recall, indicating the need for more representative training samples and clearer user guidance.

Although this study focused on one-stage detectors for real-time performance and deployment feasibility, we recognize the relevance of two-stage detectors such as Faster R-CNN in academic and high-precision contexts. In a comparative study conducted by Fatima et al. (2024), YOLOv9 outperformed Faster R-CNN in all three mAP metrics (mAP@50, mAP@75, and mAP@50-95), indicating better localization and classification precision. Moreover, YOLOv9 achieved an approximately 2.7 times faster inference speed (0.0311 s vs. 0.0844 s) [19], making it significantly more suitable for mobile or edge device deployment. These results reinforce our model selection in favor of practical deployment requirements without compromising accuracy.

4.5. Implications for Environmental Policy and Monitoring Practice

This study demonstrated the potential of combining AI and citizen science for asbestos monitoring. In scenarios in which official resources are limited, integrating AI detection with community reporting offers a viable alternative for asbestos risk management. Spatial statistical methods such as Moran’s I and LISA [20,21] can be incorporated to identify risk hotspots and guide targeted remediation efforts and policy implementation.

5. Conclusions

This study presents a practical integration framework combining deep learning (YOLOv9-E) and multi-source remote sensing imagery, including Google Street View and public participation GIS, to detect asbestos tiles affected by light occlusion. Unlike previous efforts that primarily focused on rooftops, our approach emphasized the detection of materials on the building side, addressing a critical blind spot in existing datasets.

Through social participation, citizens contributed cellular phone photos that were enriched with geolocation and orientation metadata. These user-generated images effectively complemented GSV imagery in capturing oblique and occluded views of the structures. The YOLOv9-E model, trained on this hybrid dataset, demonstrated strong performance, validating the applicability of deep learning in real-world environmental monitoring.

Future research directions include extending this framework to monitor other hazardous building materials, integrating spatial statistical tools to identify risk hotspots, and enhancing data quality through structured citizen training programs. Ultimately, this approach offers a scalable and inclusive pathway for participatory environmental surveillance and asbestos risk mitigation.