TOWARDS A PARADIGM SHIFT ON MAPPING MUDDY WATERS IN WATER RESERVOIRS WITH SENTINEL-2 USING MACHINE LEARNING

Heavy rainfall and landslides comprise rising extreme events due to climate change. Such events induce increased run-off, high-flow, flooding and erosion leading to increased sediment particles in the water. The erosion is especially high when soil is exposed in regions such as construction/mining sites, burned areas, or areas with poor agricultural/forestry practices, which in turn results in large volumes of sediment entering the water abruptly, constituting it as "muddy". Muddy water can have a series of negative implications on a lake ecosystem (e.g., altering the physicochemical regime), human infrastructure (e.g., dams, water utilities hardware), as well as the local economy (e.g., when used for recreational purposes), and for this reason its presence needs to be monitored. In this work we show that muddy water mapping can be accomplished with machine learning-based semantic segmentation constituting an extra source of sediment-laden water information. Among others, such an approach can solve issues including i) presence/absence, frequency and spatial extent ii) generalization and expansion to unknown reservoirs (assuming a curated global dataset) since the heavy calibration process is not needed iii) indication about presence of other pollutants since it act as their proxy, with applicability in the emergency management but also domain research. Our approach is based on 14 Sentinel-2 (S-2) scenes from inland/coastal waters around Greece and Finland. Atmospheric corrections are applied and compared based on spectral signatures, which adds to the notion that in classification problems it is the contrast among per-class pixel values that matter the most instead of true surface reflectance. Muddy water and non-muddy water samples are taken according to expert knowledge, S-2 scene classification layer, and NDTI combined with NDWI, and are evaluated based on their spectral signature statistics. Finally, a Random Forest model is trained, fine-tuned and evaluated using standard classification metrics. The experiments have shown that muddy water can be detected with high enough discrimination capacity, opening the door to more advanced machine learning techniques.