Using artificial neural networks to estimate missing rainfall data

Missing rainfall data from a time series or a spatial field of observations can present a serious obstacle to data analysis, modeling studies and operational forecasting in hydrology. Numerous schemes for replacing missing data have been proposed, ranging from simple weighted averages of data points that are nearby in time and space to complex statistically-based interpolation methods arid function fitting schemes. This paper presents a technique for replacing missing spatial data using a backpropagation neural network applied to concurrent data from nearby gauges. Tests performed on a sample of gauges in the Middle Atlantic region of the United States show that this technique produces results that compare favorably to simple techniques such as arithmetic and distance-weighted averages of the values from nearby gauges, and also to linear optimization methods such as regression.

Duke Scholars

Author Ana P. Barros Civil and Environmental Engineering