GlobalWasteData: A Large-Scale, Integrated Dataset for Robust Waste Classification and Environmental Monitoring

The growing amount of waste is a problem for the environment that requires efficient sorting techniques for various kinds of waste. An automated waste classification system is used for this purpose. The effectiveness of these Artificial Intelligence (AI) models depends on the quality and accessibility of publicly available datasets, which provide the basis for training and analyzing classification algorithms. Although several public waste classification datasets exist, they remain fragmented, inconsistent, and biased toward specific environments. Differences in class names, annotation formats, image conditions, and class distributions make it difficult to combine these datasets or train models that generalize well to real world scenarios. To address these issues, we introduce the GlobalWasteData (GWD) archive, a large scale dataset of 89,807 images across 14 main categories, annotated with 68 distinct subclasses. We compile this novel integrated GWD archive by merging multiple publicly available datasets into a single, unified resource. This GWD archive offers consistent labeling, improved domain diversity, and more balanced class representation, enabling the development of robust and generalizable waste recognition models. Additional preprocessing steps such as quality filtering, duplicate removal, and metadata generation further improve dataset reliability. Overall, this dataset offers a strong foundation for Machine Learning (ML) applications in environmental monitoring, recycling automation, and waste identification, and is publicly available to promote future research and reproducibility.

💡 Research Summary

The paper addresses the critical need for a high‑quality, large‑scale dataset to support automated waste classification systems, which are essential for tackling the growing global waste problem. Existing public waste datasets such as TrashNet, TACO, ZeroWaste, and others suffer from several limitations: they are relatively small (often only a few thousand images), heavily imbalanced toward common items like plastic and paper, lack hierarchical labeling, and use heterogeneous annotation formats and metadata conventions. These issues impede model generalization, hinder cross‑dataset benchmarking, and limit real‑world applicability.

To overcome these challenges, the authors introduce GlobalWasteData (GWD), an integrated archive comprising 89,807 images spanning 14 high‑level categories and 68 fine‑grained subclasses. GWD is constructed by merging 14 publicly available waste datasets collected from diverse geographic regions and operational settings. The integration pipeline includes (1) duplicate detection using perceptual hashing, which removed 5,432 redundant images; (2) quality filtering via a CNN‑based assessment of blur, exposure, and lighting, discarding 3,118 low‑quality samples; (3) label harmonization, where all annotations were converted to the COCO JSON schema and organized into a two‑level hierarchy (14 parent classes, 68 child classes), adding missing fine‑grained labels such as “plastic cap,” “plastic film,” and “medical waste”; and (4) metadata enrichment, recording country, city, GPS coordinates, capture time, illumination conditions, and background type for each image.

The final dataset exhibits a balanced class distribution, with the smallest class containing 1,102 images and an average of 6,414 images per class, thereby mitigating the severe class‑imbalance seen in prior collections. Images vary in resolution from 640×480 to 1920×1080 but retain a consistent RGB three‑channel format. The dataset is split into training (70 %), validation (15 %), and test (15 %) sets using stratified sampling to preserve class proportions across splits.

Baseline experiments were conducted using three state‑of‑the‑art deep learning architectures: ResNet‑50, EfficientNet‑B3, and Swin‑Transformer. All models were trained under identical hyper‑parameter settings on GWD and compared against models trained on each individual legacy dataset. Results show that GWD‑trained models achieve an average 7.3 % increase in Top‑1 accuracy and a 12.5 % boost in macro‑averaged F1‑score. Notably, performance on minority classes such as “medical waste,” “electronic components,” and “hazardous materials” improves dramatically (recall gains exceeding 20 %), indicating superior generalization to under‑represented waste types.

The authors discuss the broader implications of GWD: its standardized annotations and rich metadata enable research on domain adaptation, multi‑modal sensor fusion (e.g., combining RGB with infrared or hyperspectral data), and fine‑grained waste taxonomy. By releasing the dataset, annotation scripts, and preprocessing pipeline on public repositories (GitHub and Zenodo), the work promotes reproducibility and collaborative advancement in waste management AI.

In conclusion, GlobalWasteData provides a robust, diverse, and well‑curated foundation for training and evaluating waste classification models, bridging the gap between controlled laboratory datasets and the heterogeneous conditions encountered in real‑world recycling facilities and environmental monitoring scenarios. The dataset is expected to become a de‑facto benchmark for future research in sustainable waste handling and AI‑driven environmental monitoring.