"Multi-Task UAV Dataset for Floating Plastic Waste Detection in the Kashmir Region (AquaPlasticWaste 2026)"

"AquaPlasticWaste 2026 is a Multi-Task UAV Dataset for Floating Plastic Waste Detection in the Kashmir Region, comprising 5,000 UAV aerial imagery images for automated plastic waste monitoring on Dal Lake, an ecologically important Himalayan freshwater lake in Srinagar, Kashmir, India, facing severe pollution pressure. Images were captured with a consumer-grade quadrotor UAV flown at a consistent altitude across open water, shoreline, houseboat clusters, and floating garden zones, under varied lighting conditions. It is the first multi-task, polygon-annotated UAV dataset targeting plastic pollution on the surface of a Himalayan lake.All images include polygon-based instance segmentation masks rather than bounding boxes to capture the irregular shapes of floating litter (e.g., flattened bags, torn wrappers). Masks were created in CVAT and verified through a primary annotator, a second cross-checking annotator, a third adjudicator, and a final quality-control pass. Two labeling schemes are provided from the same masks: binary (plastic vs. non-plastic) and seven-class (background, bottle, cap, foam, polythene, shoe, wrapper). Annotations are exported in both COCO JSON and YOLO polygon\/segmentation formats for direct use in standard detection and segmentation frameworks.Of the 5,000 images, 4,692 (93.8%) contain plastic and 308 (6.2%) are plastic-free, with 30,166 polygon instances annotated in total. Wrapper is the most common class (35.4% of instances, present in 63.2% of images); shoe is the rarest (1.1% of instances, 4.8% of images). Mean plastic pixel coverage is 1.40% per image (median 0.55%, max 33.31%), and median instance size is 1,448 pixels, well below typical benchmark object sizes (e.g., MS-COCO) \u2014 a genuine small-object challenge. Imbalance is reported at three levels for different loss-function needs: 31.9:1 by instance count (wrapper vs. shoe), 44.8:1 by pixel area among foreground classes (polythene vs. cap), and 70.4:1 by overall pixel count between plastic and background.A single annotation source supports four downstream tasks: image classification, litter-density regression, instance segmentation, and semantic segmentation, enabling multi-task and cross-task benchmarking without repeated labeling effort.Existing waste-detection datasets such as TrashNet, TACO, UAVVaste, and TrashCan are ground-level, underwater, or single-task; none target the surface of a freshwater lake from the air. AquaPlasticWaste 2026 fills this gap by offering a realistic, class-imbalanced, small-object benchmark relevant to environmental monitoring, lake conservation planning, and edge-deployable UAV vision systems."