AerialWaste: Benchmark for Waste Detection
- AerialWaste is an annotated collection of 10,434 aerial images from Lombardy, Italy, used for binary classification of waste versus non-waste scenes.
- It integrates multi-source imagery from AGEA Orthophotos, WorldView-3, and Google Earth, presenting diverse resolutions and radiometric conditions.
- The dataset underpins deep learning-based environmental monitoring, enabling benchmarking of lightweight CNNs, transformers, and ensemble models.
The AerialWaste Dataset is a publicly released, annotated collection of aerial imagery for binary classification of waste-bearing versus non-waste scenes in the Lombardy region of Italy, assembled from AGEA Orthophotos, WorldView-3, and Google Earth imagery (Sharmily et al., 23 Aug 2025). It is used for automated landfill and waste detection, particularly in the context of illegal landfill monitoring, where high-quality public datasets are described as difficult to obtain because of security concerns. As characterized in the comparative study built around it, the dataset combines professionally curated imagery with heterogeneous spatial and radiometric conditions, making it relevant both as a practical environmental-monitoring resource and as a stress test for model robustness under source variation (Sharmily et al., 23 Aug 2025).
1. Definition, scope, and intended task
AerialWaste is presented as a large collection of 10,434 images from Lombardy, intended for deep learning-based environmental monitoring, with the core task defined as detecting images containing waste (dump sites/landfills) versus non-waste areas (Sharmily et al., 23 Aug 2025). The associated study frames illegal landfills as a hazardous environmental and public-health problem, and positions automated detection as a means of reducing the time, manpower, and resources required for manual discovery.
The dataset’s role is operational rather than merely illustrative. It is used as the basis for training, validation, and comparison of multiple deep learning architectures, including lightweight CNNs, transformer-based models, and ensembles. In that sense, AerialWaste functions simultaneously as a labeled corpus, an evaluation benchmark, and a reproducibility substrate for future landfill-detection work (Sharmily et al., 23 Aug 2025).
A notable point is that the paper also reports 11,703 images extracted during preprocessing from the original JSON-format dataset. This creates a documented discrepancy with the 10,434-image figure reported in the abstract. A plausible explanation is that the larger figure reflects preprocessing-stage extraction and balancing operations rather than the base released corpus, but the paper summary preserves both counts rather than collapsing them into a single reconciled total (Sharmily et al., 23 Aug 2025).
2. Composition and image provenance
The dataset aggregates imagery from three distinct sources, each with different acquisition characteristics:
| Aspect | Description |
|---|---|
| Region | Lombardy, Italy |
| Base image count | 10,434 images |
| Image sources | AGEA Orthophotos, WorldView-3, Google Earth |
The use of AGEA Orthophotos contributes high-resolution aerial photography associated with agricultural and land-use applications. WorldView-3 introduces very high-resolution commercial satellite imagery, while Google Earth contributes composite, multi-sensor, variable-quality imagery (Sharmily et al., 23 Aug 2025). This heterogeneity is not incidental; the paper explicitly notes that aerial images from these sources exhibit variability in resolution and quality, which can hamper consistent performance across datasets.
The dataset is nonetheless described as professionally curated, diverse and high-quality, and these two properties—curation quality on the one hand and source heterogeneity on the other—define much of its technical significance (Sharmily et al., 23 Aug 2025). The former supports supervised learning reliability; the latter introduces realistic nuisance variation. For model development, this means that performance on AerialWaste depends not only on class discrimination but also on tolerance to changes in resolution, rendering, and imaging conditions.
3. Annotation schema, curation, and preprocessing
AerialWaste includes metadata and annotations provided in JSON files, with each image labeled as either containing waste (positive) or not containing waste (negative) (Sharmily et al., 23 Aug 2025). The paper further reports a manual review stage in which samples were relabeled based on manual inspection to correct annotation errors, including moving clear cases from negative to positive and vice versa. This makes annotation correction part of the dataset’s effective construction rather than a peripheral cleanup step.
Preprocessing standardizes images to 256×256 pixels or, for some models, 224×224 pixels (Sharmily et al., 23 Aug 2025). The study also applies explicit normalization using per-channel statistics:
input_mean = [0.3201, 0.3334, 0.2832]input_std = [0.2004, 0.1818, 0.1764]
Class imbalance is reported in the original data, with more negatives than positives. To address this, the positive class was augmented using rotation, flipping, color jitter, and cropping. The same augmentation family is described as serving both balancing and generalization purposes (Sharmily et al., 23 Aug 2025). Because the dataset combines multiple image sources with inherent quality differences, color jitter is especially notable: it is used to simulate lighting variation and image-quality differences rather than merely to enlarge the training set.
These preprocessing and curation stages imply that AerialWaste is not just a static image repository. It is a curated classification corpus whose practical use depends on label correction, harmonized resizing, normalization, and imbalance mitigation. This suggests that downstream conclusions about model behavior on the dataset are partly conditioned by the quality-control pipeline, not only by the raw imagery itself.
4. Partitioning, training use, and evaluation protocol
The paper reports an initial split with 7,278 images for training-related use and 2,607 images for testing (Sharmily et al., 23 Aug 2025). Of the 7,278 images, 20% (1,818 images) were reserved for validation, leaving 5,460 images as the training set. After augmentation for class balancing, the training folders for both positive and negative classes are reported as containing 4,852 images each.
The dataset is used to train and validate a range of architectures:
- Lightweight CNNs: MobileNetV2, SqueezeNet, DenseNet121, GoogLeNet
- Transformer-based and hybrid models: MobileViT-XS, ViT-Tiny
- Ensemble approaches: Parallel Ensemble (MobileViT-XS + ViT-Tiny), and three-model fusion (Sharmily et al., 23 Aug 2025)
The study uses Negative Log-Likelihood Loss (NLLLoss) with log-softmax model outputs. Performance is measured using confusion-matrix-derived quantities—Accuracy, Precision, Sensitivity (Recall), F1-score, and Specificity—and also visualized through ROC analysis with AUC (Sharmily et al., 23 Aug 2025). The reported metric definitions are:
The paper also states that ablation studies on optimizers found AdamW to be the most robust for this task (Sharmily et al., 23 Aug 2025). Within the dataset’s experimental role, AerialWaste therefore supports both architectural comparison and training-strategy comparison.
5. Empirical behavior and benchmark outcomes
The principal empirical conclusion drawn from experiments on AerialWaste is that complex and heavy models were found to be more prone to overfitting and memorizing training data instead of learning patterns, whereas lightweight simpler models achieved significant success with less overfitting (Sharmily et al., 23 Aug 2025). This observation is central to how the dataset is interpreted: despite being large by the standards of public landfill imagery, it does not necessarily reward model scale.
The best reported result is obtained through an ensemble and fusion technique, yielding the following weighted-average metrics for binary classification:
| Model/result | Metrics |
|---|---|
| Best fusion model | Accuracy 92.33%, Precision 92.67%, Sensitivity 92.33%, F1-score 92.41%, Specificity 92.71% |
| Parallel Ensemble | Accuracy 91.56%, F1-score 91.59% |
| MobileNetV2_050, 10 frozen layers | Accuracy 90.68%, F1-score 90.69% |
The study also reports AUC = 0.97 for the best ensemble model’s ROC analysis (Sharmily et al., 23 Aug 2025). Comparative experiments identify MobileNetV2, SqueezeNet, and related lightweight architectures as especially effective relative to deeper CNNs such as DenseNet and GoogLeNet, with overfitting cited as the likely cause of the deeper models’ weaker behavior.
Within the paper’s documentation, this evidence is distributed across Table III (per-model performance), Table IV (parallel ensemble versus single models), Table V (confusion-matrix metrics per class and weighted average), and Figure 1 (ROC curve) (Sharmily et al., 23 Aug 2025). The dataset thus serves not merely as a source of labels but as the empirical basis for a broader architectural claim: in this domain, model parsimony may generalize better than model depth.
6. Significance, limitations, and relation to adjacent aerial-imagery datasets
AerialWaste is significant because it addresses a domain where good quality publicly released datasets for illegal landfill detection are hard to find due to security concerns (Sharmily et al., 23 Aug 2025). Its public availability, multi-source composition, and manually corrected annotations make it a practical benchmark for reproducible research in waste detection and environmental monitoring.
Its limitations are also explicit. The dataset begins with class imbalance, exhibits variability in resolution and quality, and contains annotation errors that required correction (Sharmily et al., 23 Aug 2025). These properties complicate interpretation of model performance. The paper’s own results suggest that source heterogeneity and limited positive-sample structure can induce overfitting in deeper architectures, even after augmentation and balancing. This suggests that reported performance should be read as contingent on both dataset design and model-selection strategy.
In the broader literature, AerialWaste occupies a distinct niche among aerial and UAV-based environmental datasets. A later open-access waste-detection study targets openly dumped dispersed municipal solid waste in crowdsourced UAV imagery across 29 regions in 10 countries in Sub-Saharan Africa, using 5 m × 5 m tiles and reporting 92.87% accuracy on its test set (Knoblauch et al., 4 May 2026). That work emphasizes cross-site generalization in UAV contexts, whereas AerialWaste, as described here, is a multi-source image-level binary classification corpus centered on Lombardy. By contrast, the Leafy Spurge Dataset addresses invasive plant detection in drone imagery rather than waste detection, with 500 presence and 500 absence samples from 10 sites and a best reported 0.84 accuracy using a pre-trained DINOv2 vision transformer (Doherty et al., 2024). Together, these datasets indicate that aerial environmental monitoring is increasingly organized around public benchmarks, but they differ substantially in label granularity, sensor provenance, and target semantics.
For landfill detection specifically, AerialWaste’s enduring importance lies in the conjunction of three properties: public accessibility, source diversity, and documented benchmark use (Sharmily et al., 23 Aug 2025). Those characteristics make it useful not only for training classifiers but also for studying annotation quality, domain shift, model efficiency, and ensemble behavior in real-world remote-sensing pipelines.