Can satellite based pattern-oriented memory improve the interpolation of sparse historical rainfall records?

- New approach to the interpolation of historical sparse rainfall data sets.
- Broad basis for the comparison of the newly proposed method.
- Very promising results of the newly proposed pattern-oriented memory method.
- Pattern-oriented memory models show great margin for improvement.
- One of the first uses of Least Squares Support Vector Regression for hydrological applications.

SummaryThere is a standing challenge in obtaining long localized records of rainfall data in many large river basins of the developing world. Recent spaceborne instrumentation offers a consistent source of rainfall information, but this information covers only a relatively limited time period. In this context, and given its consistence, a question rises on the potential offered by this new wealth of information to improve our understanding of the rainfall patterns and how to use them in order to alleviate the historical problems of scarcity of observed historical records.The present research focuses on the interpolation of historical rainfall records over large spatial scales and low availability of observed point data, with distances between measurement points in the order of tenths to hundreds of kilometers and temporal scales ranging from daily to monthly. The main goals of the work are twofold: firstly, to evaluate the potential of using a novel pattern-oriented interpolation technique to learn complex spatial rainfall patterns from satellite data and applying this knowledge in the interpolation of historical rainfall maps; secondly, to assess the performance of the proposed methodology by comparing its results to those of other interpolation techniques suitable for spatially sparse datasets.The proposed pattern-oriented interpolation technique uses modern data sources to enhance the reliability of the interpolation of historical rainfall areal distributions. Results show that, under given conditions, the pattern-oriented memory class of models can considerably reduce the errors traditionally associated with historical rainfall interpolation at large spatial scales and under low availability of spatial data.