View Proposal

Proposer
Ian Tan
Title
Storm Chaser: From Short-Range Tracking to Extended 6-Hour Tropical Rainfall Nowcasting
Goal
To reproduce a baseline deep learning nowcasting in the 1st semester, establish a working benchmark for short-range storm tracking (0–2 hours), and then extend the project scope to evaluate model degradation over an extended forecast horizon (3–6 hours) in tropical convective climates.
Description
Predicting high-intensity (severe) rainfall in tropical climates is uniquely challenging. Short-range forecasts (0.5 - 2 hours) are dominated by storm movement (advection), whereas extended-range forecasts (typically 3–6 hours) require predicting the rapid initiation, evolution, and decay of convective storm cells. This is important as allowing more time for authorities to react to severe storms will ensure safety of the citizens. For this project, the expected progress would be: Phase 1: Reproduction and Baseline Setup (September Semester) Reproduce an established work by Shi et al. (2015) (or similar) using Convolutional LSTM. Train using radar sequence snapshot sequences. You will need to set up the data pre-processing pipeline, train the model for nowcasting (0.5 - 2 hours), and determine/compute the baseline metrics using measures such as MSE. Phase 2: Extend the Horizon (January Semester) Building on what was done in Phase 1, you are to scale the output sequence to 3 - 6 hours using possibly a modified U-Net architecture for longer temporal sequence (or EarthFormer by Gao et al.). The focus for this phase 2 is to analyse where and why the model struggles over longer lead times especially in the areas under consideration (spatial blurness). Potential Dataset: Hong Kong, being an area where early warning on severe weather is common will serve as a good initial dataset for the benchmarking of the 0.5 - 2 hour models. This can be obtained from HKO-7 (Hong Kong Observatory Radar Dataset), where it captures high-intensity monsoon and subtropical convective storm sequences. SEVIR (Storm EVent ImagRy), a benchmark dataset from Lincoln Laboratory (MIT) featuring heavy convective storm events. The dataset includes nicely aligned multi-sensor channels, allowing you to experiment with multi-modal inputs for phase 2. MyMETData by Jabatan Meteorologi Malaysia has image and radar based data for weather that can be accessed (https://mymetdata.met.gov.my/shop/category/met-data/4). There is a fee involved and depending on the progress of the project, this may be a good source for the project.
Resources
A single consumer or cloud GPU (e.g., NVIDIA 3080 or Google Colab T4 will be sufficient). Do note that the training time will be in hours (unsure yet) and maybe up to a day for Phase 2. Some references: Shi, X., et al. (2015). "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting." NeurIPS. Gao, Z., et al. (2022). "Earthformer: Exploring Space-Time Transformers for Earth System Forecasting." NeurIPS.
Background
Url
Difficulty Level
Moderate
Ethical Approval
None
Number Of Students
0
Supervisor
Ian Tan
Keywords
precipitation nowcasting, radar extrapolation, convolutional lstm, tropical convective storms, u-net
Degrees
Bachelor of Science in Computing Science
Back to List