Ensemble deep learning with image captioning for visual pollution detection, classification, and reporting.

With the rapid urban development and initiatives such as Saudi Vision 2030, efforts have been directed toward improving services and quality of life in Saudi cities. As a result, multiple environmental challenges have emerged, including visual pollution (VP), which significantly impacts the quality of life. Current approaches to these challenges rely on reporting through an online application managed by the Ministry of Municipalities and Housing, which is prone to errors due to manual data entry. This study proposes an AI-driven framework that integrates deep learning models (YOLOv5 and EfficientDet), along with ensemble techniques. Additionally, the study proposes using Bootstrapping Language-Image Pre-training (BLIP) to automatically generate text descriptions based on the content of images in reports. This framework was developed using the public dataset "Saudi Arabia Public Roads Visual Pollution Dataset" from Mendeley. This study is the first to combine the results of the YOLOv5 and EfficientDet models to detect VP and automatically generate descriptions using BLIP-2, thereby facilitating the production of citizen-monitored reports. The proposed system aims to improve decision-making, reduce errors, and enhance urban management by automating the detection, classification, and reporting of VP. This ensemble approach achieved a Mean Average Precision (mAP) of 0.95, a recall of 0.95, a precision of 0.91, and an F1 score of 0.93, surpassing the performance of the individual models. In image captioning, the "BLIP2-Flan-T5-XL" model achieved an accuracy 80% based on human evaluation, demonstrating the effectiveness of AI-generated text in urban reporting. This suggests that the system could help automate VP reporting and improve reporting accuracy, thereby contributing to more sustainable cities.