Search

In-depth evaluation and analysis of the error properties associated with satellite-based precipitation estimation algorithms can play an important role in the future development and improvements of these products. This study evaluates the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) daily data set from 2000 to 2011 in 69 pixels over a semiarid basin in northwest Iran and compares it with the data set of the existing rain-gauge network. Different analytical approaches and measures are used to examine PERSIANN performance seasonally and categorically. The residuals are also decomposed into true positive (hit), false negative (miss), and false alarm (FA) estimate biases in addition to systematic and random error components. The results show seasonal variability of PERSIANN precision in rainfall detection with substantial errors during winter and summer that are associated with high rates of FA ratio (more than 60%). The value of miss and FA biases (124 and -77,000mm, respectively, within the total data set) are considerably larger than hit and total bias (27 and 74,000mm, respectively) because these components contribute conversely and compensate each other by their opposite signs. Moreover, PERSIANN detects heavy rainfalls well with a probability of detection (POD) over 80%, but with serious biases. Generally, although the detection ability of PERSIANN improves as the rate of rainfall increases, its systematic error in simulation of the rainfall process also increases (from 5% systematic error to 90% in heavier rainfalls), leading to a low level of accuracy in the estimation of precipitation rate.