Quality control procedures for sub-hourly rainfall data: An investigation in different spatio-temporal scales in Brazil

Abstract

• 1st automatic quality control on sub-hourly scale applied to Brazil rainfall dataset. • An automatic procedure that does not use a reference quality-controlled dataset. • Rainfall properties were used to spatially identify malfunctioning gauges. • 93.6% of high-quality and 78.6% of poor-quality gauges were correctly identified. There has been an increase in the intensity of rainfall extreme events. These events have a high contribution to hazards in urban areas, such as flash flooding, and landslides, which lead to social, economic, and environmental damage. Sub-hourly rainfall information has a critical role in the assessment of such natural disasters. However, precipitation gauge networks in developing countries still lack high-quality data. This study presents a new automatic quality control procedure (A-QCP) for sub-hourly rainfall data in Brazil, analysing 7 years (2014–2020) of tipping bucket rain gauge (TBRG) data from over 3,000 gauges per year provided by CEMADEN’s (Brazilian National Centre for Monitoring and Early Warnings of Natural Disasters) rain gauge network as a case study. The A-QCP developed separately examines each year through a series of single-gauge tests in which each station was considered independently, going through procedures to assess possible equipment malfunctions through data analysis like long rainless periods or with constant tips due to clogging, spurious rainfall peaks, and long periods of missing data. Then, nearest neighbour analysis was performed through spatial outlier detection using Local Moran’s Index. The performance of the method was evaluated through a confusion matrix, which compared the procedures’ results with a reference dataset created using visual and independent analyses. After the application of the proposed A-QCP to the whole database, the results show that the procedure was able to correctly identify 78.6 % of faulty rain gauges for the whole period, with less than 8 % of functioning rain gauges removed from the final High-Quality database. It was also noted that the best performing filters were the first ones applied, showing that the procedure works best on the identification of gross errors, with diminished efficiency in finding malfunctions after the sequential removal of Poor-Quality rain gauges.