Automatic water features recognition from Satellite images using an Artificial Neural Network

Abstract

Water body is a significant natural object and 70% of total Earth is covered with water. Out of which approx. 68% is saline water and is part of many water resources such as glaciers, oceans, sea, ponds etc. and only 2% is fit for drinking or portable water. Water body is an area that have well defined topographical boundary where the water accumulates, for example: river, ocean, sea, lake, reservoir etc. Mapping of water bodies is important in flood prediction, environmental monitoring, safe navigation, environmental protection, Geographical Information System (GIS) database updating, sustainable development & planning, watershed definition, evaluation of water resources and many more. Traditionally, mapping of water bodies in small areas are carried out using conventional field surveying methods or water bodies are manually delineated with a pencil on vellum paper overlaid on top of aerial photographs or traced with a cursor from digital remote sensing images on the computer screen. These methods are tedious, subjective and time-consuming. Automatic feature extraction is a beneficial method to get updated spatial data from aerial images, satellite images and Digital Surface Model (DSM) instead of traditional methods. Therefore, an effective technique is tested for automatic water features extraction from satellite images based on Artificial Neural Network (ANN). Basically, the methodology has two stages as learning and application. In learning stage, defined ANN is trained using a small subset of the satellite image in study area and numbers of subset are simulated in application stage. Then, a shape file of the vector layer of extracted water bodies is provided automatically as the final output of the system. The methodology is tested for worldview-02 satellite images in the ?Samanalawewa? reservoir and the surrounding area of Belihuloya, Sri Lanka. The system is provided the accuracy as average completeness: 98.97%, correctness: 98.27% and quality: 97.28%.