The final objective of this project is to generate depth distribution for each cell mesh given different preipitation criteria.

Flow Design

Depth estimator training

Train a regression transformer to estimate the depth by 592 real events involving channel, terrain, precipitation-based features and region-based features (optional).

cCTGAN modeling

Random sample 50 real events to train CTGAN with features: x, y, cumu_rain, peak_int and duration.

Add constraints:

• Positive constraints
• Inequalty constraints: cumu_rain $\geq$ peak_int
• Custimized logic: peak_int $\geq$ cumu_rain / duration Save checkpoints for each pair of D/G learning rates and every 50 epochs and generate 5,000,00 samples each for grid search.

Synthetic depth generation

For the optimal checkpoint, concatenate the synthetic precipitation-based features and cooresponding spatial features to predict synthetic depth by trained depth estimator.

After that, we will form separated synthetic events by following methods.

Synthetic events separation Design

Real events distribution determination

For each cell:

• Get distributions of cumu_rain, peak_int, and duration.
• Determine Levels: [Low, Medium, High] for each distribution by:
  • Low: $[0, mean - coef_1 \cdot std)$
  • Medium: $(mean - coef_1 \cdot std, mean + coef_2 \cdot std]$
  • High: $> mean + coef_2 \cdot std$

Synthetic events distribution control

For each cell:

• Separate each distribution into 3 segments by determined real thresholds
• Check if rows filling in the overlapping of the determined class among 3 distributions
  • Yes: Move rows into the determined class pool.
  • No: Drop.
• When creating determined class events, random sample the rows from their corresponding pools.
  • *: if the pool size of specific rows is lower than the average size, add slight noise when sampling.

Synthetic events generation

• Given input command Low, Medium or High, event_duration will be sampled randomly from the pool and set globally and then the corresponding event_cumu_rain, event_peak_int will be sampled randomly from the pool as well given the constriants of event_duration.
• The sampled synthetic precipitation-based features will concatenate with corresonding spatial features channel and terrain to form a complete feature input.
• To address the NULL caused by feature pool selection, each NULL will be the average value of its k nearest neighbors (KNN).
• In addition, KNN smoother will be applied on event_cumu_rain and event_peak_int.
• Finally, the synthetic depth generated by depth estimator will concatenate with features to plot maps.