WASTE DUMP IDENTIFICATION BY SEMANTIC SEGMENTATION USING DEEP LEARNING AND SENTINEL 2

Increase in total amount of waste produced at global level has serious ecosystem implications, from health risks for individuals to environmental damage. This increase is directly correlated with the urban development in the past decades, increase in consumption of goods and services or in large industrial activities. In EMERITUS project, we aim to support waste management activities with the help of Earth Observation (EO) driven products derived from Sentinel 2 using deep learning segmentation approaches. To achieve a segmentation model that can properly identify waste dumps, where training datasets are missing, manual labelling and various models testing was implied. Using tools such as Optuna, a more straightforward approach is used for hyperparameters search. It is observed that the proposed model identifies waste dumps locations in different Sentinel 2 tiles, with main false positives in areas with very high spectral mixing (e.g. parking lots).