Go to Paper
Return to GeoComputation 99 Index

Multi-Model Data Fusion for Hydrological Forecasting

SEE, Linda (L.See@geog.leeds.ac.uk) ABRAHART, Robert J. (babrahart@hotmail.com), Center for Computational Geography, School of Geography, University of Leeds, Leeds, LS2 9JT, U.K.

Key Words: data fusion, neural networks, hydrological modelling

Data fusion is an expanding area of research that deals with the integration of information from multiple sensors and/or data sources to obtain a better solution than could otherwise be achieved with the use of single-source data on its own. The different approaches that could be adopted for fusing multisource data include conventional, statistical, and numerical methods as well as artificial intelligence techniques such as neural networks, genetic algorithms, and fuzzy-logic. The emergence of new sensors, advanced fusion algorithms, together with improved hardware and software, make real-time data fusion a practical option for automated target recognition systems, applied robotics, and remote sensing. The opportunities that are afforded from data fusion also are important in other areas of commerce and science, and are seen to have clear implications for different types of geographical research and in application-related fields such as hydrological modelling. Data fusion can be implemented at different scales of operation and with different strengths and complexities. The simplest type of data fusion operation would involve the integration of input data from multiple sources to produce output data. This is referred to as the "data in - data out" fusion architecture. But data fusion also can operate at more complicated or elaborate feature-based or decision-based levels which could, in turn, have significant implications for the design and construction of enhanced automated flood prediction and flood warning systems.

This paper provides results from some initial explorations that have been undertaken on the use of neural networks for fusing hydrological river flow forecasts, and involves a comparison between two different river catchments in the United Kingdom-the Upper River Wye and the River Ouse. The data sources in this instance were not sensors but different predictive models that have been run in parallel. The operation was based on fusing multisource model output data, obtained from conventional statistical and numerical methods, and from various artificial intelligence techniques. The new data fusion predictions were based on different combinations of these single-model predictions, with some additional experiments being carried out in association with the historical hydrological record.