The 3D Neural Network for improving radar-rainfall estimation in monsoon climate
Weather radar can offer synoptic measurement at a higher temporal and spatial resolution to extract the rain information. Rainfall can be inverted from the radar reflectivity using the power-law relation to ground rain gauge measurement. The relationship known as Z-R model has been established in many variants but the uncertainty from the sampling bias and the Z-R variability of single-polarization radar observation on monsoon rain becomes subject to research. The inconsistence of the Z-R model exemplifies the main disadvantages of radar rainfall estimation. Previously proposed universal Z-R models are inadequate to be applied for heterogenous rain intensity during monsoon seasons . Studies on the accuracy of Z-R model are lacking for many reasons such as complex radar data structure, requires a large data storage and is very expensive when used for long-term observations. Radar reflectivity data was collected by the S-band single-horizontal polarization Doppler radar at Kota Bahru, Kelantan and the data is managed by Malaysia Meteorological Department (MMD). The observation covers radial measurement from 50 to 250 km radii. The reflectivity data represents the return signal sampled in 10 minutes and at position within 2-km altitude from the ground which was generated from the lowest scanning angle of 0.7 degree. Such setup would reduce the ground clutter effect which is typically pronounced in the signal received close to the radar antenna providing reliable rainfall information near to the surface and also minimizing the temporal uncertainty in the Constant Plan Position Indicator (CAPPI). Calibration through the threshold log receiver of signal-to-noise ratio (LOG) and the Doppler channel clutter-to-signal ratio (CSR) were applied in pre-processing for minimizing the uncertainties formed by the beam blockage, ground clutter, and during instrument calibration (VAISALA 2016). The radar data conversion from SIGMET to Netcdf was applied through Python Atmospheric Radiation Measurement (ARM) Radar Toolkit (Py-ART). Tipping bucket rain gauges provide rain intensity at a volumetric resolution of 0.2 mm in 1-hour resolution, organised by the Department of Irrigation and Drainage (DID). For this study, there are 58 gauges involved in the 200-km radar observation range. These studies reported that 50 gauges have experienced more than 10% of data void and 41 gauges were working in homogenous fashion while the rest of gauges were doubtful. The digital elevation model (DEM) data is a product of The Shuttle Radar Topography Mission (SRTM) with 30 meters (1 arc-second) pixel spacing and was used for altitude correction of each rain gauge. The DEM products have been corrected on the void pixels they can be downloaded from the U.S. Geological Survey (USGS) website. Reflectivity and rain gauge data are obtained during the wet season of the northeast monsoon from 2013 to 2015 (January, February, March, September, October, November and December).
Steps to reproduce
Two column data consist of the matching pair of the rain gauge (R) and the reflectivity (Z), had been corrected for the altitude correction.The the was extracted at corresponding gauge location in hourly basis. The rainfall was estimated by using parametric Z-R model known as LM and ANN. For LM, the coefficients obtained from the non-linear least square regression optimization was applied, followed by the spatial bias adjustment. Finally, bias factor, B is typically applied on the derived Z-R model with the prefactor alpha and coefficient Beta. Apply the NN function to estimate the rainfall for all intensity and different intensity.. For spatial rainfall pattern, the NN function and LM for all intensity was used.