Data Mining of Remotely-Sensed Rainfall for a Large-Scale Rain Gauge Network Design

Rain gauges are able to measure point rainfall with high accuracy compared with remote sensing observations. However, a single rain gauge cannot provide continuous spatial coverage, and thus, rain gauge networks need to be designed in a way that will provide optimum rainfall information with a minimum number of gauges. While relatively inaccurate but long-term larger-scale satellite rainfall measurements are an ideal dataset to provide insight into local storm characteristics, such as rainfall patterns and local topographic influences on rainfall, these key characteristics would be useful in designing rain gauge networks. This article proposes a new scheme for a large-scale rain gauge network design that uses complementary data from satellite rainfall measurements. Principal component analysis (PCA), the elbow method, the intra-group sum of squares error and the gap statistic were used to calculate the optimum number of rain gauges respectively, while the locations of rain gauges were determined by cluster analysis with the influence of rainfall amount. The results show that PCA is the most effective method with multilevel variances, providing an optimum reference design number at 80% variances. In addition, the rainfall zones in Kenya had a close relationship with land cover, and more gauges will have to be deployed in the mountainous areas to reflect rainfall variations during the warm season in Kenya. The proposed scheme can also be used for other types of ground observation station designs in conjunction with remotely sensed hydrometeorological factors such as soil moisture, evapotranspiration and ice cover.