Computer Vision with Explainable Artificial Intelligence for Visual Pollution Detection in the Kingdom of Saudi Arabia

• Published in: IEEE access 2024-12, p.1-1
• Main Authors: Mazroa, Alanoud Al, Maray, Mohammed, Alashjaee, Abdullah M., Alotaibi, Faiz Abdullah, Alzahrani, Ahmad A., Alkharashi, Abdulwhab, Alotaibi, Shoayee Dlaim, Alnfiai, Mrim M.
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Computer Vision; Contrast Enhancement; Explainable AI; Explainable Artificial Intelligence; Noise; Object recognition; Optimization; Pollution; Prairie Dog Optimization; Roads; Tuning; Urban areas; Visual Pollution Detection; Visualization
• ISSN: 2169-3536
• DOI: 10.1109/ACCESS.2024.3513696

Environmental pollution often results from numerous human actions. Researchers have studied the risks and impacts of harmful pollutants and environmental contamination for years, leading to the implementation of several critical measures. New solutions are continuously advanced to tackle this problem effectively. Visual pollution extends outside advertising, demonstrated in numerous forms through natural areas, urban, and roadways. Among the plethora of various procedures of visual pollution, environmental pollution worsens the aesthetics of the city, approving the significance of investigation and evaluating it from multiple dimensions. Building automated pollutants or pollution detection methods became progressively popular owing to the present growth of improved artificial intelligence methods. While some developments are made, automatic pollution detection must still be fully understood and well-researched. Therefore, this study focuses on designing and developing the Modeling of Computer Vision with Explainable Artificial Intelligence for Visual Pollution Detection (MCVXAI-VPD) model. The MCVXAI-VPD model involves DL-based object detection and classification with a hyperparameter tuning strategy. In the developed MCVXAI-VPD methodology, an original pre-processing stage occurs in two levels: mean filter (MF)-based noise removal and CLAHE-based contrast enhancement. Next, the MCVXAI-VPD model applies a YOLOv5 object detector with a backbone network combination of CSP and SPP to effectually detect the target objects. Besides, the MCVXAI-VPD model performs a classification process using deep learning depending on bidirectional long short-term memory (BiLSTM). Additionally, the prairie dog optimization (PDO) technique is exploited as a hyperparameter tuning process of the BiLSTM model to accomplish enhanced classification performance. At last, the MCVXAI-VPD methods integrate the XAI model LIME to enhance the explainability and understanding of the black-box technique, ensuring more accurate detection of VP. A comprehensive experimental study has been performed to ensure the improved performance of the MCVXAI-VPD method. The performance validation of the MCVXAI-VPD method portrayed a superior accuracy value of 98.20% over existing techniques.