Validation of Volunteered Geographic Information on Urban Visual Pollution Using Deep Learning Algorithms

ABSTRACT
Visual pollution is one of the most significant challenges in urban landscape management. Traditional monitoring methods, despite their high accuracy, are time-consuming, costly, and limited in spatial coverage. In contrast, citizen-centered spatial data enable broader and more cost-effective monitoring; however, the heterogeneity and potential errors in crowdsourced data highlight the need for their scientific validation. This study aims to validate citizen-centered spatial data in assessing visual pollution on urban walls and to develop a hybrid framework combining deep learning and citizen participation, thereby enhancing the reliability of these data for urban decision-making.The study employed a mixed-methods approach in four stages. First, spatial and visual data on visual pollution were collected from various areas of Tehran via a web-based platform. Next, the images were labeled into four classes. In the third stage, three convolutional neural network (CNN) models—ResNet50, EfficientNetB0, and EfficientNetV2-L—were trained. Finally, by comparing model outputs with citizen labels, a validation mechanism for the data was developed and model performance was evaluated.The EfficientNetV2-L model achieved the highest accuracy at 87.78% and showed greater stability in classifying difficult data. Learning curves confirmed stable convergence and effective control of overfitting.The results demonstrate that integrating citizen-centered data with deep learning models provides an efficient and reliable framework for monitoring visual pollution. This framework can serve as a dynamic reference for validating spatial data and as an effective tool for intelligent urban landscape management.
 
Extended Abstract
Introduction
Visual pollution, as one of the growing challenges in contemporary cities, exerts a multilayered and profound impact on the quality of the urban environment, the mental well-being of citizens, and the way public spaces are perceived. The urban landscape, as the setting that citizens encounter on a daily basis, plays a significant role in shaping feelings of belonging, safety, satisfaction, and place identity. When this landscape becomes visually disordered, its consequences extend beyond a mere decline in aesthetic quality and may lead to heightened feelings of abandonment, disorder, reduced public trust, and the weakening of social capital. In this context, urban walls, due to their spatial extent, continuous visibility, and capacity to function as carriers of both formal and informal messages, play a substantial role in shaping the visual quality of the city.
Urban walls can perform a dual function. On the one hand, they may serve as platforms for urban art, convey local identity, and strengthen a sense of belonging. On the other hand, in the absence of effective management and supervision, they can become sites for the accumulation of illegal advertisements, deteriorated posters, wall writings, unauthorized graffiti, and signs of physical decay. This condition is particularly intensified in high-density areas and disadvantaged neighborhoods, contributing to the reproduction of spatial inequality across the city. Accordingly, understanding the extent of visual pollution on urban walls and continuously monitoring it are fundamental prerequisites for urban landscape management.
Despite the importance of this issue, conventional approaches to monitoring visual pollution have largely relied on field observations, expert assessments, and qualitative judgments. While these methods can yield acceptable results in limited projects and small-scale contexts, they face serious limitations at the metropolitan scale. The high cost of field studies, the time-consuming nature of data collection, dependence on specialized human resources, and the lack of rapid repeatability are among the main challenges associated with these approaches. Moreover, the outcomes of such studies are often static and cross-sectional, limiting their capacity to capture temporal and spatial changes in a timely manner.
In recent years, citizen-generated data have attracted increasing attention as a novel source of urban information. Citizen participation in reporting and documenting environmental conditions enables the production of large volumes of data with extensive spatial and temporal coverage. In addition to reducing monitoring costs, this approach enhances social participation, increases public awareness, and strengthens civic responsibility toward the living environment. Nevertheless, participatory data are inherently heterogeneous and uncertain. Variations in users’ levels of knowledge, differences in perceptual interpretations of concepts, poor image quality, and spatial inaccuracies are among the factors that may challenge the reliability of such data.
At the same time, recent advances in deep learning and urban image analysis have created new opportunities for automated and scalable monitoring of visual phenomena. Convolutional neural networks demonstrate a strong capacity to extract visual features and classify images, enabling them to identify complex patterns that are difficult or time-consuming for humans to detect. However, these models are also highly dependent on accurate and reliable training data, and the use of noisy or biased data may lead to errors or systematic distortions. Consequently, the central issue addressed by this research lies at the intersection of these two domains: how to leverage the potential of citizen-generated data and the analytical power of deep learning within a complementary, self-correcting framework simultaneously.
The primary objective of this study is to design and test a human–machine feedback mechanism to validate citizen-generated data on visual pollution on urban walls. Within this framework, humans and machines are not treated as independent sources of information, but rather as components of a learning system that mutually enhance one another’s performance. By focusing on urban walls as one of the most prominent elements of the urban landscape, the study enables a more precise definition of indicators and reduces conceptual ambiguity surrounding visual pollution.
From a conceptual perspective, visual pollution in this research is defined through a multidimensional framework encompassing physical, perceptual, and behavioral dimensions. The physical dimension includes manifestations such as illegal advertisements, deteriorated posters, dilapidated and abandoned walls. The perceptual dimension addresses citizens’ interpretations of order, aesthetics, and environmental quality, while the behavioral dimension examines how these conditions influence social actions and reactions. This framework allows for a clearer distinction between visually disruptive walls and those possessing artistic or cultural value.
 
Methodology
The research adopts a mixed approach, integrating field data, citizen participation, and deep learning algorithms. Tehran was selected as the study area due to its large scale, high degree of morphological diversity, and significant levels of visual pollution, providing an appropriate context for testing the proposed framework. The selection of this city enabled the examination of a wide range of urban wall conditions and enhanced the generalizability of the findings. The data collection process followed two complementary paths. In the first path, researchers conducted systematic field surveys to capture images of urban walls. These images were collected to establish a controlled reference dataset and played a crucial role in the initial training of deep learning models. Attention was given to controlling image quality, viewing angles, and spatial accuracy to minimize data errors. In the second path, a web-based platform was designed and implemented to facilitate citizen participation. This platform allowed citizens to upload images of urban walls along with their geographic locations, enabling voluntary reporting of observed conditions. This stage significantly increased the volume of data and improved the spatial coverage of the study, while also promoting active citizen engagement in the monitoring process. Following data collection, a data cleaning process was conducted to enhance dataset quality. Duplicate, low-quality, irrelevant, or spatially inaccurate images were removed. This initial filtering was essential to reduce noise and prevent the negative influence of unreliable data on model training. Subsequently, the remaining images were labeled according to predefined visual pollution categories. The categories used in the study included walls with illegal advertisements, wall writings and unauthorized graffiti, deteriorated and abandoned walls, and clean or artistically painted walls as a control class. Label quality was ensured through random reviews and corrections of ambiguous cases, significantly reducing human labeling errors and improving the reliability of the training data. After preparing the dataset, several convolutional neural network architectures were employed for image classification. Images were preprocessed and augmented before being divided into training, validation, and test sets. Models were trained with consistent settings to enable fair performance comparisons.  A range of evaluation metrics was applied to enable a detailed assessment of each model’s strengths and weaknesses.
 
Results and discussion
The results of model training and evaluation indicated that deeper, better-optimized models achieved superior performance in identifying different manifestations of visual pollution. These models were particularly effective in distinguishing subtle differences between walls with artistic value and those that constituted visual disturbance. In contrast, simpler models exhibited higher error rates in certain categories, highlighting the importance of selecting appropriate architectures for analyzing participatory data. One of the most significant components of the study is the validation of citizen-generated data through the human–machine feedback mechanism. In this process, the label assigned by the citizen is compared with the output generated by the deep learning model. When the two labels are consistent, the data point is considered reliable. In cases of inconsistency, the data are flagged as requiring review and can be reintroduced into a correction cycle. This process enables the gradual elimination of unreliable data and improves the overall quality of the dataset. Simultaneously, the deep learning model benefits from receiving corrected data, thereby enhancing its performance over time. As a result, a dynamic feedback loop is established that concurrently improves data quality and model accuracy. Analysis of the results reveals that the highest level of agreement between citizen labels and model predictions occurs in clearer categories, such as clean walls or walls with illegal advertisements. Conversely, the greatest discrepancies are observed in distinguishing unauthorized graffiti from artistic wall paintings. This finding reflects the perceptual complexity of these categories and underscores the influence of cultural and social context on citizens’ interpretations. Overall, the findings demonstrate that citizen-generated data, when supported by deep learning models and appropriate validation mechanisms, can serve as a reliable source for monitoring urban visual pollution. The proposed framework enables the production of timely information, identification of visual pollution hotspots, and prioritization of urban management interventions.
 
 
Conclusion
In conclusion, the study demonstrates that integrating citizen-generated data with deep learning offers a novel, cost-effective, and scalable approach to monitoring visual pollution. Despite limitations such as the focus on a single city and specific visual categories, the proposed framework shows strong potential for extension and application in future research. It can serve as a strategic tool for urban landscape management and the enhancement of environmental quality in cities.
 
Funding
There is no funding support.
 
Authors’ Contribution
Authors contributed equally to the conceptualization and writing of the article. All of the authors approved thecontent of the manuscript and agreed on all aspects of the work declaration of competing interest none.
 
Conflict of Interest
Authors declared no conflict of interest.
 
Acknowledgments
We are grateful to all the scientific consultants of this paper.