DeepFlood | Hai Dang Trinh

linkOrigin

DeepFlood began as a team project at SEAS. After the program, I continued the project independently, refining the data pipeline, validation design, peak-sensitive training strategy, baseline comparison, and evaluation dashboard. This led to a more rigorous research question:

Can a basin-specific time-series model preserve rare local streamflow extremes without leaking future information into validation?

The result is a same-day streamflow nowcasting prototype specific to the Long Đại basin, Quảng Bình, Vietnam.

linkProblem

Extreme hydrological events create an imbalanced regression problem. Most observations represent normal flow conditions, while the most important events are rare peaks.

Time-series experiments also require careful validation design because random splits and preprocessing that uses full-dataset statistics can introduce future information into historical evaluation. This can inflate apparent performance without reflecting real predictive skill.

linkData & Method

Each prediction uses a seven-day input window of hydrometeorological observations. The model estimates streamflow on day $t$ using meteorological observations available through day $t$ and streamflow observations available before day $t$ .

The features derive from three stages of processing:

Raw observations

Precipitation (prcp) · Max/min temperature (tmax, tmin) · Solar radiation (srad) · Vapor pressure (vp) · Day length (dayl) · Historical streamflow

↓

Engineered signals

Rolling rainfall windows · Seasonality indicators · Lagged streamflow · Temporal statistics

↓

Model input

39 features × 7-day window · target-derived streamflow features use strictly lagged observations

X_t \in \mathbb{R}^{7 \times 39}

where 7 = temporal sequence length (days) and 39 = engineered hydrometeorological features

The sequence is processed through a sequential architecture:

X_t \rightarrow \mathrm{Conv1D}_{64} \rightarrow \mathrm{BiLSTM}_{128} \rightarrow \mathrm{BiLSTM}_{64} \rightarrow \mathrm{Temporal\ Attention} \rightarrow \hat{Q}_t

The bidirectional layers operate only within the observed input window. They do not access observations after prediction time $t$ , and target-derived streamflow features use historical lags only.

The continued work focused on controlling future data leakage risk, refining feature generation, enforcing chronological validation, testing peak-sensitive weighting, and evaluating reproducible artifacts.

To penalize large flow errors during training, samples are weighted by flow magnitude:

w_i = 1 + 30\left(\frac{y_i}{y_{\max,\mathrm{train}}}\right)

1 \le w_i \le 31

within the training target range

Here $y_i$ is the training target for sample $i$ and $y_{\max,\mathrm{train}}$ is the maximum streamflow in the training period. High-flow observations become progressively more influential during optimization. This places greater emphasis on large flow errors, but it may also increase average absolute error on ordinary observations. The current experiment does not isolate its individual contribution.

linkValidation

To prevent temporal leakage, evaluation enforces a strict chronological split:

EARLIERLATER

Training period · 70%

Validation period · 30%

Validation period contains 107 chronological observations. The same period was used for checkpoint selection.

Safeguard	What it prevents
Chronological split	Future observations being randomly mixed into the training set
Train-only scaling	Future distribution information entering preprocessing scalers
Strictly lagged streamflow	Day-T target information leaking into day-T model input features

linkResults

Performance comparison across the 107-observation validation period:

Metric	DeepFlood	Persistence baseline
MAE	130.8 m³/s	126.9 m³/s
RMSE	148.1 m³/s	312.8 m³/s
NSE	0.726	−0.222

DeepFlood reduces RMSE by approximately 52.6% relative to persistence. DeepFlood MAE is approximately 3.0% higher than persistence.

Diagnostic on the largest event in the validation period:

Validation peak diagnostic	Value
Observed peak	2,576.39 m³/s
Predicted peak	2,800.68 m³/s
Magnitude error	+8.7%
Timing lag	0 days

This diagnostic covers a single event and is not evidence of reliable peak performance across extreme events.

linkInterpretation

DeepFlood does not outperform persistence on every metric. Persistence achieves slightly lower MAE because streamflow often changes gradually between consecutive days, so the persistence assumption remains competitive under normal conditions.

DeepFlood substantially improves RMSE and NSE. This reflects the intended trade-off: peak-sensitive training places more weight on large flow errors during optimization, which emphasizes high flow deviations at the cost of a small increase in average absolute error on ordinary observations. The metrics emphasize different error behavior. MAE treats all errors equally, while RMSE penalizes large deviations more heavily.

linkLimitations

Evidence supports	Evidence does NOT establish
Same-day streamflow nowcasting for Long Đại	Multi-day forecasting skill
Chronological validation period performance	Performance on an independent test set
Basin-specific results	Cross-basin generalization
Precomputed evaluation dashboard	Operational warning deployment
Peak-sensitive behavior during validation period	Reliability across all extreme events
Comparison against a persistence baseline	Superiority over simpler ML models or alternative architectures

linkArtifact

Evaluation dashboard: https://deepflood.haidangtrih.me/

Predictions shown are precomputed evaluation outputs. The dashboard visualizes hindcast results and is not an operational inference service.

Teaching a flood model to hear one river.