Automated Detection and Mapping of Waste in Different Environments and Seasons Using UAV Imagery and AI
Description
This presentation describes a workflow for automated detection and mapping of waste in different environments using UAV imagery and artificial intelligence. The study focuses on two sites in western Călărași County, Romania, near the Argeș River and the localities of Budești and Șoldanu. RGB UAV imagery acquired during the spring 2026 campaign is processed using semantic segmentation with a DeepLabV3+ model and a ResNet-152 encoder. The model considers four waste categories: plastic bags, plastic bottles, construction debris, and rubber. Predictions are converted into spatial summaries, including object counts and approximate mapped area, using image geolocation and ground sampling distance estimates. Preliminary post-processing tests examine confidence thresholds and overlap-controlled frame selection to reduce repeated counting. At a confidence threshold of 0.5, overlap control reduced the reported object count from 521 to 194 and summed mapped area from 37.02 m² to 12.59 m². Seasonal robustness and generalisation across sites remain to be evaluated, as only the spring campaign has been completed at this stage. The work is associated with the ACCELERATE project (101182930), funded by the European Union under Horizon Europe Marie Skłodowska-Curie Actions Staff Exchanges.
Files
Files
(7.4 MB)
| Name | Size | |
|---|---|---|
|
md5:94409c81f448bc9eec6faaf3c0f8a706
|
7.4 MB | Download |