Road cleanliness is a vital aspect of city management and hygiene, but nearly always traditional methods of monitoring using manual methods are costly and inefficient. The current study explores the feasibility of deep-learning models for the automated recognition of road cleanliness, using a publicly available Kaggle corpus of 237 images, which are categorized according to clean and dirty. Six pretrained Convolutional Neural Network (CNN) models were fine-tuned and compared to a Vision Transformer (ViT), MobileNetV2, ResNet50, DenseNet121, InceptionV3, VGG16, and VGG19. Before the model was trained, the images were resized spatially, normalized in intensity, and data augmented to increase generalization. The model performance was measured in terms of accuracy, precision, recall, and F1 score. VGG19 became the most accurate CNN with 96%, and ResNet50 the least accurate with 79% accuracy. The ViT has the better score, where the general accuracy was 98% with equal precision and recall. Local Interpretable Model-Agnostic Explanations (LIME) were to be used to permit easy interpretation, as they showed that salient road-surface features dominated the prediction of the model. These findings support the feasibility of ViT, as well as explainable artificial intelligence (XAI) methods, to provide high-quality and verifiable evaluations for environmental monitoring.