| Summary: | Water reservoirs are an essential resource for human health, natural ecosystems, and socio-economic activities, making their effective monitoring mandatory for informed decision-making on sustainable water management. Particularly important is monitoring the extent and level, which enable determination of volume variations. In this respect, this work investigates the performance of Segment Anything Model (SAM) – a foundation model for segmentation released by Meta AI Research – in segmenting water bodies from medium-resolution satellite imagery. SAM was applied in its original form to Sentinel-1 and Sentinel-2 images through prompt engineering (seed modality), testing five different 3-band combinations of the input images in two areas of the Ligurian coast (Italy). Overall, the study demonstrates the adaptability and efficiency of SAM in segmenting water bodies. Among the tested configurations, the SAR 3-band combination achieved the best performance, with [Formula: see text] scores ranging from 0.874 to 0.994. Furthermore, SAM performs better in simpler scenarios with uniform radiometric properties and regular water boundaries, achieving [Formula: see text] score close to 1 with seeds in any position within the water body. Conversely, in more complex scenarios, accurate seed prompt placement becomes critical; analysis of [Formula: see text] maps revealed that seeds placed near the edge of the water, where sharp gradients occur, significantly improve SAM segmentation performance.
|
|---|
