Assessing the impact of dataset quality on the performance of artificial intelligence models in automatic waste classification

The growing importance of sustainable waste management fosters the development of artificial intelligence-based solutions for automatic waste classification. However, the effectiveness of these systems largely depends on the quality of the training data. The aim of this study was to analyse publicly available datasets in terms of their representativeness, class balance, and usefulness under real-world sorting conditions. Two datasets of differing complexity - TrashNet and WaRP - were used in the experiments. The results show that datasets with a uniform structure, such as TrashNet, enable high classification accuracy, but may lead to overfitting to the training data, which results in decreased performance on the validation set. Using this dataset, our models achieved the highest F1-scores for cardboard (0.84), paper (0.75), and glass (0.64), with an overall accuracy of 0.66. Realistic datasets, such as WaRP, better reflect actual conditions but pose challenges related to class imbalance and high visual variability. When trained on WaRP, the models performed best on detergent boxes (F1-score: 0.73) and various milk bottle categories (F1- scores ranging from 0.64 to 0.66). The study highlights the importance of well-prepared datasets in designing reliable AI systems for waste classification.