A New Self-Learning Algorithm for Dynamic Classification of Water Bodies

In many applications of remote sensing data land-water masks play an important role. In this context they can be a helpful orientation to distinguish dark areas (e.g. cloud shadows, topographic shadows, burned areas, coniferous forests) and water areas. However, water bodies cannot always be classified exactly on basis of available remote sensing data. This fact can be caused by a variety of different physical and biological factors (e.g. chlorophyll, suspended particles, surface roughness, turbid and shallow water and dynamic of water bodies) as well as atmospheric factors (e.g. haze and clouds). On the other hand the best available static water masks also show deficiencies. These are essentially caused by the fact that land-water masks represent only a temporal snapshot of the water bodies distributed worldwide and therefore these masks cannot reflect their dynamic behavior. This paper presents a dynamic selflearning water masking approach for AATSR remote sensing data in the context of integrating high-quality water masks in processing chains for deriving valueadded remote sensing data products. As an advantage to conventional water masking algorithms, the proposed approach operates on basis of a static water mask as data base for deriving an optimized dynamic water mask. Significant research effort was spent to develop and validate a dynamic self-learning algorithm and a processing scheme for operational derivation of actual landwater masks as basis for operational interpretation of remote sensing data. Based on this concept actual activities and perspectives for contributions to operational monitoring systems will be presented.