"AquaSurf-Malnad-223: A Vegetation-Stratified Benchmark for False-Positive Analysis in Freshwater Debris Detection"
Year: 2026
Type: dataset
Abstract: "Automated debris detection on inland freshwater is usually evaluated on coastal or open-water imagery, where the water surface is largely unobstructed. AquaSurf-Malnad-223 targets the opposite regime: 223 geotagged photographs of vegetation-dense freshwater bodies in the Malnad region of Karnataka, India, captured in a single session in April 2025. Each image carries an ordinal rating of surface-vegetation coverage, and every bounding box was inspected individually as a cropped image and either confirmed, relabelled, or rejected, yielding 85 verified debris annotations over a four-class solid-waste taxonomy together with 197 images confirmed to contain no debris.The dataset is released as a precision and false-positive benchmark rather than a large-scale training corpus. A COCO-pretrained Faster R-CNN evaluated on it under a uniform protocol achieves 5.0% proposal precision overall, and precision degrades monotonically with surface-vegetation coverage, from 38.5% on lightly vegetated water to 1.4% where floating hyacinth dominates the frame. The verified negatives make this false-positive rate measurable, which is not possible in corpora containing only positive frames. The dataset supports domain-shift evaluation, false-positive analysis and vegetation-aware model selection for inland water monitoring."
Source: IEEE DataPort
DOI: 10.21227/k8d4-5n64
Authors: Shaziya Banu S, Pallavi G.B
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7169727307