Thesis Code: T7

Bias Correction and Spatial Adjustment of Global Extreme Rainfall Events in Reanalysis Data Using Deep Learning

Short description:

Hydro-meteorological reanalysis datasets, such as ERA5 [1] produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), are key resources for studying historical weather events. Indeed, precipitation fields from ERA5 are often used as inputs to hydrological models and risk assessment analyses of flood and compound impacts, as well as to provide the initial conditions for real-time operational hydrological forecasts and reforecasts. However, studies have found that ERA5 systematically underestimates heavy precipitation events and have suggested that ERA5 rainfall data must be bias corrected before being used in hydrological models [2]. Most existing works on bias adjustment focus on adjusting pixel-wise metrics of bias, such as the Mean Squared Error (MSE). However, it is equally important to ensure that the rainfall peaks are correctly located within rainfall maps, especially if these maps are then used as input to hydrological models.

In the past year, members of the eiLab have developed a deep learning algorithm for adjusting such biases, focusing on rainfall maps of tropical cyclones. The algorithm uses an off-the-shelf Convolutional Neural Network (CNN) and an innovative loss function, which is able to substantially reduce the biases of ERA5 both at the pixel level and in terms of spatial patterns of extreme rainfall.

The goal of this thesis is to extend this work to all extreme rainfall events in ERA5, re-training the algorithm on a larger and more varied set of events and adjusting it as necessary, for instance by adding new predictors to the network. As tropical cyclones tracks are mostly over the oceans, it was not possible in the previous work to include data such as orography. For generic extreme rainfall events over land, it is expected that such inputs would greatly aid the network.

An application of the thesis will be to simulate the flood extent driven by historical extreme rainfall events using the adjusted rainfall reanalysis as input to an hydrological and hydrodynamic model. For this the SFINCS model developed at Deltares will be used, which is a state-of-the-art process-based model that can simulate the compound effects of multiple flood drivers (rainfall, storm surge and river flooding) and has shown great potential in achieving high accuracy at limited computational costs [5]. Simulated flood extents will be validated against observed flooded areas from satellite images [6]. This will allow us to assess the impact of the meteorological data adjustment on the accuracy of flood simulations, which are key inputs for informing climate adaptation measures, such as designing flood-related insurance schemes or planning structural defences (e.g., levees, dykes, etc.).

To achieve these goals, the student will carry out the following activities: 

  • Literature review: review the state of the art for rainfall bias adjustment algorithms and extreme precipitation events.
  • Data elaboration:
    • Identify extreme rainfall events in ERA5 using a detection algorithm called ‘Forward-in-time’, developed by Skok et al. (2009, 2010) [3-4] and available off-the-shelf.
    • Extract ERA5 data corresponding to the extreme rainfall events identified, including variables such as total precipitation, total cloud cover, orography.
  • Computational experiments:
    • Re-train the available CNN on the new dataset.
    • Test different combinations of inputs, and potentially implement slight modifications to the model to optimise it further.
    • Simulate flood events using a hydrodynamic model (SFINCS) and assess the flood extent simulation accuracy using observed flood extent from satellite data.

References

  1. Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Horányi, A., Muñoz‐Sabater, J., … & Thépaut, J. N. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999-2049
  2. Ansari, R., & Grossi, G. (2022). Performance evaluation of raw and bias-corrected ERA5 precipitation data with respect to extreme precipitation analysis: case study in Upper Jhelum Basin, South Asia. Theoretical and Applied Climatology, 150(3-4), 1409-1424
  3. Skok, G., J. Tribbia, J. Rakovec, and B. Brown (2009), Object-based analysis of satellite-derived precipitation systems over the low- and midlatitude Pacific Ocean, Mon. Weather Rev., 137(10), 3196–3218, doi:10.1175/2009MWR2900.1
  4. Skok, G., J. Tribbia, and J. Rakovec (2010), Object-based analysis and verification of WRF model precipitation in the low- and midlatitude Pacific Ocean, Mon. Weather Rev., 138(12), 4561–4575, doi:10.1175/2010MWR3472.1
  5. Leijnse, T., van Ormondt, M., Nederhoff, K., & van Dongeren, A. (2021). Modeling compound flooding in coastal systems using a computationally efficient reduced-physics solver: Including fluvial, pluvial, tidal, wind- and wave-driven processes. Coastal Engineering, 163, 103796. https://doi.org/10.1016/j.coastaleng.2020.103796
  6. Tellman, B., Sullivan, J. A., Kuhn, C., Kettner, A. J., Doyle, C. S., Brakenridge, G. R., Erickson, T. A., & Slayback, D. A. (2021). Satellite imaging reveals increased proportion of population exposed to floods. Nature, 596(7870), 80–86. https://doi.org/10.1038/s41586-021-03695-w

Relevant courses and knowledge:

Natural Resources Management

Number of students:

1

Requisites:

The student must be comfortable with coding (Python will be used for the project). Knowledge of machine learning tools is a major plus.