<p>Solar energy is one of the most rapidly growing sources of renewable energy, making the inspection and cleaning of photovoltaic panels essential for maintaining efficiency. Accumulation of dust and other environmental factors degrade panel performance over time. This study conducts a comparative analysis of multiple machine learning models for PV module classification based on soiling detection, employing a standardized preprocessing approach and a locally acquired dataset. The dataset consists of 2158 images captured from residential PV panels in Karachi, Pakistan, under real-world conditions. Various preprocessing techniques, including data augmentation, were applied to train and evaluate machine learning models, including CNN, ResNet, DenseNet, YOLOv8n, and the latest YOLOv11n. Models initially trained with limited epochs exhibited moderate performance, with YOLOv11n achieving the highest accuracy of 94.5%, outperforming YOLOv8n (84.0%), CNN (91.4%), DenseNet (81.9%), and ResNet (77.0%). Extending the training to optimal epochs with advanced preprocessing and data augmentation significantly improved classification performance. YOLOv11 achieved an exceptional 99.0% accuracy, precision, recall, and F1 score, making it the most reliable model for soiling detection. These findings demonstrate that extended training with advanced preprocessing significantly enhances accuracy, with YOLOv11 emerging as the most effective model for real-world PV soiling detection. This study introduces a new custom-labeled image dataset captured under natural environmental conditions and conducts one of the first comprehensive comparative analyses involving YOLOv11 and other deep learning models for PV soiling detection.</p>