Deep Learning Precipitation Downscaling
M.Sc. Thesis Project | Central Department of Hydrology and Meteorology, Tribhuvan University
2024 | M.Sc. Thesis Grant (Equivalent to 2300 USD)
Project Overview
This research develops and implements deep learning algorithms to downscale coarse-resolution ERA5 precipitation data to high-resolution grids suitable for local-scale hydrometeorological applications. The work focuses on Nepal's Roshi Khola Basin, a mountainous region with complex terrain and limited ground-based precipitation observations.
The project bridges the gap between global climate datasets and local hydrological modeling needs, enabling more accurate flood risk assessment and water resource management in data-sparse regions.
Methodology
Deep Learning Architecture: U-Net
Implemented a U-Net convolutional neural network framework, originally developed for image segmentation, adapted for precipitation downscaling. The architecture captures multi-scale spatial features essential for accurate precipitation representation over complex mountainous terrain.
Input Data Processing
- Source Dataset: ERA5 reanalysis precipitation at 0.25° resolution (~28 km)
- Target Resolution: 3 km grid spacing for high-resolution hydrometeorological applications
- Domain: Roshi Khola Basin, Nepal
- Temporal Coverage: Multiple years of daily precipitation data
- Auxiliary Inputs: Topographic features (elevation, terrain slope) to capture orographic effects
Validation Strategy
Comprehensive performance evaluation using three independent precipitation datasets:
- ERA5-Land: High-resolution reanalysis product (0.1° resolution)
- IMERG: Satellite-based precipitation estimates (0.1° resolution)
- Station Observations: Ground-truth measurements from meteorological stations
Evaluation conducted at daily timescales using standard metrics (RMSE, MAE, correlation coefficient, bias).
Hydrological Application
Implemented an empirical flood discharge estimation model at hourly timescales using downscaled precipitation as input. This enables generation of runoff forecasts and flood risk assessment with improved spatial resolution.
Key Results & Outcomes
Resolution Enhancement
Successfully downscaled coarse ERA5 data from 28 km (~0.25°) to 3 km resolution, enabling better representation of local precipitation variability and orographic effects.
Model Performance
Downscaled precipitation dataset demonstrates improved accuracy compared to native ERA5 when validated against ERA5-Land, IMERG, and station observations.
Hydrological Validation
Empirical flood discharge model utilizing downscaled precipitation inputs produces realistic runoff forecasts at hourly timescales.
Practical Application
Framework is transferable to other mountainous basins in South Asia with limited ground-based observations.
Technologies & Tools
Programming
- Python
- TensorFlow / Keras
Data Processing
- xarray / netCDF
- NumPy / Pandas
- Rasterio
Climate Data
- ERA5 Reanalysis
- IMERG Satellite
- Station Observations
Hydrological Modeling
- Empirical Discharge Models
- Runoff Estimation
- Flood Forecasting
Research Impact & Significance
This project addresses a critical challenge in climate science and water resources management: the availability of high-resolution precipitation data in mountainous regions. By combining deep learning with domain knowledge, the research enables:
- Improved Flood Risk Assessment: High-resolution precipitation enables more accurate flood forecasting for flood-prone communities.
- Water Resource Management: Better understanding of spatial precipitation patterns for hydropower planning and irrigation management.
- Climate Adaptation: Supports decision-making for climate change adaptation strategies in water-dependent sectors.
- Open Science: Framework and methodology can be applied to other data-sparse regions globally.
Future Work & Extensions
- Expand the framework to other basins across Nepal with varying topographic and climatic conditions
- Incorporate additional climate variables (temperature, humidity) for multi-variate downscaling
- Develop probabilistic downscaling methods to quantify uncertainty
- Apply to future climate projections for flood risk assessment under climate change scenarios