Institut für Erd- und Umweltwissenschaften
Refine
Year of publication
Document Type
- Article (19)
- Postprint (7)
- Habilitation Thesis (1)
- Other (1)
Language
- English (28)
Is part of the Bibliography
- yes (28)
Keywords
- Algorithm (2)
- Band (2)
- Kwajalein (2)
- Methodology (2)
- Reflectivity (2)
- Uncertainties (2)
- Weather (2)
- models (2)
- skill (2)
- weather (2)
Institute
RainNet v1.0
(2020)
In this study, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. Its design was inspired by the U-Net and SegNet families of deep learning models, which were originally designed for binary segmentation tasks. RainNet was trained to predict continuous precipitation intensities at a lead time of 5min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of 900km × 900km and has a resolution of 1km in space and 5min in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In order to achieve a lead time of 1h, a recursive approach was implemented by using RainNet predictions at 5min lead times as model inputs for longer lead times. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the rainymotion library and had previously been shown to outperform DWD's operational nowcasting model for the same set of verification events.
RainNet significantly outperforms the benchmark models at all lead times up to 60min for the routine verification metrics mean absolute error (MAE) and the critical success index (CSI) at intensity thresholds of 0.125, 1, and 5mm h⁻¹. However, rainymotion turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15mm h⁻¹). The limited ability of RainNet to predict heavy rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5min, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16km and below. Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5min lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5min, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance in terms of a binary segmentation task. Furthermore, we suggest additional input data that could help to better identify situations with imminent precipitation dynamics. The model code, pretrained weights, and training data are provided in open repositories as an input for such future studies.
RainNet v1.0
(2020)
In this study, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. Its design was inspired by the U-Net and SegNet families of deep learning models, which were originally designed for binary segmentation tasks. RainNet was trained to predict continuous precipitation intensities at a lead time of 5min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of 900km × 900km and has a resolution of 1km in space and 5min in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In order to achieve a lead time of 1h, a recursive approach was implemented by using RainNet predictions at 5min lead times as model inputs for longer lead times. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the rainymotion library and had previously been shown to outperform DWD's operational nowcasting model for the same set of verification events.
RainNet significantly outperforms the benchmark models at all lead times up to 60min for the routine verification metrics mean absolute error (MAE) and the critical success index (CSI) at intensity thresholds of 0.125, 1, and 5mm h⁻¹. However, rainymotion turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15mm h⁻¹). The limited ability of RainNet to predict heavy rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5min, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16km and below. Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5min lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5min, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance in terms of a binary segmentation task. Furthermore, we suggest additional input data that could help to better identify situations with imminent precipitation dynamics. The model code, pretrained weights, and training data are provided in open repositories as an input for such future studies.
In 2009, a group of prominent Earth scientists introduced the "planetary boundaries" (PB) framework: they suggested nine global control variables, and defined corresponding "thresholds which, if crossed, could generate unacceptable environmental change". The concept builds on systems theory, and views Earth as a complex adaptive system in which anthropogenic disturbances may trigger nonlinear, abrupt, and irreversible changes at the global scale, and "push the Earth system outside the stable environmental state of the Holocene". While the idea has been remarkably successful in both science and policy circles, it has also raised fundamental concerns, as the majority of suggested processes and their corresponding planetary boundaries do not operate at the global scale, and thus apparently lack the potential to trigger abrupt planetary changes. This paper picks up the debate with specific regard to the planetary boundary on "global freshwater use". While the bio-physical impacts of excessive water consumption are typically confined to the river basin scale, the PB proponents argue that water-induced environmental disasters could build up to planetary-scale feedbacks and system failures. So far, however, no evidence has been presented to corroborate that hypothesis. Furthermore, no coherent approach has been presented to what extent a planetary threshold value could reflect the risk of regional environmental disaster. To be sure, the PB framework was revised in 2015, extending the planetary freshwater boundary with a set of basin-level boundaries inferred from environmental water flow assumptions. Yet, no new evidence was presented, either with respect to the ability of those basin-level boundaries to reflect the risk of regional regime shifts or with respect to a potential mechanism linking river basins to the planetary scale. So while the idea of a planetary boundary on freshwater use appears intriguing, the line of arguments presented so far remains speculative and implicatory. As long as Earth system science does not present compelling evidence, the exercise of assigning actual numbers to such a boundary is arbitrary, premature, and misleading. Taken as a basis for waterrelated policy and management decisions, though, the idea transforms from misleading to dangerous, as it implies that we can globally offset water-related environmental impacts. A planetary boundary on freshwater use should thus be disapproved and actively refuted by the hydrological and water resources community.
High precipitation quantiles tend to rise with temperature, following the so-called Clausius-Clapeyron (CC) scaling. It is often reported that the CC-scaling relation breaks down and even reverts for very high temperatures. In our study, we investigate this reversal using observational climate data from 142 stations across Germany. One of the suggested meteorological explanations for the breakdown is limited moisture supply. Here we argue that, instead, it could simply originate from undersampling. As rainfall frequency generally decreases with higher temperatures, rainfall intensities as dictated by CC scaling are less likely to be recorded than for moderate temperatures. Empirical quantiles are conventionally estimated from order statistics via various forms of plotting position formulas. They have in common that their largest representable return period is given by the sample size. In small samples, high quantiles are underestimated accordingly. The small-sample effect is weaker, or disappears completely, when using parametric quantile estimates from a generalized Pareto distribution (GPD) fitted with L moments. For those, we obtain quantiles of rainfall intensities that continue to rise with temperature.
Many institutions struggle to tap into the potential of their large archives of radar reflectivity: these data are often affected by miscalibration, yet the bias is typically unknown and temporally volatile. Still, relative calibration techniques can be used to correct the measurements a posteriori. For that purpose, the usage of spaceborne reflectivity observations from the Tropical Rainfall Measuring Mission (TRMM) and Global Precipitation Measurement (GPM) platforms has become increasingly popular: the calibration bias of a ground radar (GR) is estimated from its average reflectivity difference to the spaceborne radar (SR). Recently, Crisologo et al. (2018) introduced a formal procedure to enhance the reliability of such estimates: each match between SR and GR observations is assigned a quality index, and the calibration bias is inferred as a quality-weighted average of the differences between SR and GR. The relevance of quality was exemplified for the Subic S-band radar in the Philippines, which is greatly affected by partial beam blockage.
The present study extends the concept of quality-weighted averaging by accounting for path-integrated attenuation (PIA) in addition to beam blockage. This extension becomes vital for radars that operate at the C or X band. Correspondingly, the study setup includes a C-band radar that substantially overlaps with the S-band radar. Based on the extended quality-weighting approach, we retrieve, for each of the two ground radars, a time series of calibration bias estimates from suitable SR overpasses. As a result of applying these estimates to correct the ground radar observations, the consistency between the ground radars in the region of overlap increased substantially. Furthermore, we investigated if the bias estimates can be interpolated in time, so that ground radar observations can be corrected even in the absence of prompt SR overpasses. We found that a moving average approach was most suitable for that purpose, although limited by the absence of explicit records of radar maintenance operations.
Many institutions struggle to tap into the potential of their large archives of radar reflectivity: these data are often affected by miscalibration, yet the bias is typically unknown and temporally volatile. Still, relative calibration techniques can be used to correct the measurements a posteriori. For that purpose, the usage of spaceborne reflectivity observations from the Tropical Rainfall Measuring Mission (TRMM) and Global Precipitation Measurement (GPM) platforms has become increasingly popular: the calibration bias of a ground radar (GR) is estimated from its average reflectivity difference to the spaceborne radar (SR). Recently, Crisologo et al. (2018) introduced a formal procedure to enhance the reliability of such estimates: each match between SR and GR observations is assigned a quality index, and the calibration bias is inferred as a quality-weighted average of the differences between SR and GR. The relevance of quality was exemplified for the Subic S-band radar in the Philippines, which is greatly affected by partial beam blockage.
The present study extends the concept of quality-weighted averaging by accounting for path-integrated attenuation (PIA) in addition to beam blockage. This extension becomes vital for radars that operate at the C or X band. Correspondingly, the study setup includes a C-band radar that substantially overlaps with the S-band radar. Based on the extended quality-weighting approach, we retrieve, for each of the two ground radars, a time series of calibration bias estimates from suitable SR overpasses. As a result of applying these estimates to correct the ground radar observations, the consistency between the ground radars in the region of overlap increased substantially. Furthermore, we investigated if the bias estimates can be interpolated in time, so that ground radar observations can be corrected even in the absence of prompt SR overpasses. We found that a moving average approach was most suitable for that purpose, although limited by the absence of explicit records of radar maintenance operations.
This case study evaluates the suitability of radar-based quantitative precipitation estimates (QPEs) for the simulation of streamflow in the Marikina River Basin (MRB), the Philippines. Hourly radar-based QPEs were produced from reflectivity that had been observed by an S-band radar located about 90 km from the MRB. Radar data processing and precipitation estimation were carried out using the open source library wradlib. To assess the added value of the radar-based QPE, we used spatially interpolated rain gauge observations (gauge-only (GO) product) as a benchmark. Rain gauge observations were also used to quantify rainfall estimation errors at the point scale. At the point scale, the radar-based QPE outperformed the GO product in 2012, while for 2013, the performance was similar. For both periods, estimation errors substantially increased from daily to the hourly accumulation intervals. Despite this fact, both rainfall estimation methods allowed for a good representation of observed streamflow when used to force a hydrological simulation model of the MRB. Furthermore, the results of the hydrological simulation were consistent with rainfall verification at the point scale: the radar-based QPE performed better than the GO product in 2012, and equivalently in 2013. Altogether, we could demonstrate that, in terms of streamflow simulation, the radar-based QPE can perform as good as or even better than the GO product - even for a basin such as the MRB which has a comparatively dense rain gauge network. This suggests good prospects for using radar-based QPE to simulate and forecast streamflow in other parts of the Philippines where rain gauge networks are not as dense.
Rainfall-induced attenuation is a major source of underestimation for radar-based precipitation estimation at C-band. Unconstrained gate-by-gate correction procedures are known to be inherently unstable and thus not suited for unsupervised attenuation correction. In this study, we evaluate three different procedures to constrain gate-by-gate attenuation correction using reflectivity as the only input. These procedures are benchmarked against rainfall estimates from uncorrected radar data, using six years of radar observations from the single-polarized C-band radar in South-West Germany. The precipitation estimation error is obtained by comparing the radar-based estimates to rain gauge observations. All attenuation correction procedures benchmarked in this study lead to an effective improvement of precipitation estimation. The first method caps the corrections if the rain intensity increase exceeds a factor of two. The second method decreases the parameters of the attenuation correction iteratively for every radar beam calculation until attaining a stability criterion. The second method outperforms the first method and leads to a consistent distribution of path-integrated attenuation along the radar beam. As a third method, we propose a slight modification of Kraemer's approach which allows users to exert better control over attenuation correction by introducing an additional constraint that prevents unplausible corrections in cases of dramatic signal losses.
Storm runoff from the Marikina River Basin frequently causes flood events in the Philippine capital region Metro Manila. This paper presents and evaluates a system to predict short-term runoff from the upper part of that basin (380km(2)). It was designed as a possible component of an operational warning system yet to be installed. For the purpose of forecast verification, hindcasts of streamflow were generated for a period of 15 months with a time-continuous, conceptual hydrological model. The latter was fed with real-time observations of rainfall. Both ground observations and weather radar data were tested as rainfall forcings. The radar-based precipitation estimates clearly outperformed the raingauge-based estimates in the hydrological verification. Nevertheless, the quality of the deterministic short-term runoff forecasts was found to be limited. For the radar-based predictions, the reduction of variance for lead times of 1, 2 and 3hours was 0.61, 0.62 and 0.54, respectively, with reference to a no-forecast scenario, i.e. persistence. The probability of detection for major increases in streamflow was typically less than 0.5. Given the significance of flood events in the Marikina Basin, more effort needs to be put into the reduction of forecast errors and the quantification of remaining uncertainties.
We explore the potential of spaceborne radar (SR) observations from the Ku-band precipitation radars onboard the Tropical Rainfall Measuring Mission (TRMM) and Global Precipitation Measurement (GPM) satellites as a reference to quantify the ground radar (GR) reflectivity bias. To this end, the 3-D volume-matching algorithm proposed by Schwaller and Morris (2011) is implemented and applied to 5 years (2012–2016) of observations. We further extend the procedure by a framework to take into account the data quality of each ground radar bin. Through these methods, we are able to assign a quality index to each matching SR–GR volume, and thus compute the GR calibration bias as a quality-weighted average of reflectivity differences in any sample of matching GR–SR volumes. We exemplify the idea of quality-weighted averaging by using the beam blockage fraction as the basis of a quality index. As a result, we can increase the consistency of SR and GR observations, and thus the precision of calibration bias estimates. The remaining scatter between GR and SR reflectivity as well as the variability of bias estimates between overpass events indicate, however, that other error sources are not yet fully addressed. Still, our study provides a framework to introduce any other quality variables that are considered relevant in a specific context. The code that implements our analysis is based on the wradlib open-source software library, and is, together with the data, publicly available to monitor radar calibration or to scrutinize long series of archived radar data back to December 1997, when TRMM became operational.