Evaluation of random forests for short-term daily streamflow forecasting in rainfall- and snowmelt-driven watersheds

Leo Triet Pham, Lifeng Luo, Andrew O. FinleyHydrology 2021 · DOI

Abstract

In the past decades, data-driven machine-learning (ML) models have emerged as promising tools for short-term streamflow forecasting. Among other qualities, the popularity of ML models for such applications is due to their relative ease in implementation, less strict distributional assumption, and competitive computational and predictive performance. Despite the encouraging results, most applications of ML for streamflow forecasting have been limited to watersheds in which rainfall is the major source of runoff. In this study, we evaluate the potential of random forests (RFs), a popular ML method, to make streamflow forecasts at 1 d of lead time at 86 watersheds in the Pacific Northwest. These watersheds cover diverse climatic conditions and physiographic settings and exhibit varied contributions of rainfall and snowmelt to their streamflow. Watersheds are classified into three hydrologic regimes based on the timing of center-of-annual flow volume: rainfall-dominated, transient, and snowmelt-dominated. RF performance is benchmarked against naïve and multiple linear regression (MLR) models and evaluated using four criteria: coefficient of determination, root mean squared error, mean absolute error, and Kling–Gupta efficiency (KGE). Model evaluation scores suggest that the RF performs better in snowmelt-driven watersheds compared to rainfall-driven watersheds. The largest improvements in forecasts compared to benchmark models are found among rainfall-driven watersheds. RF performance deteriorates with increases in catchment slope and soil sandiness. We note disagreement between two popular measures of RF variable importance and recommend jointly considering these measures with the physical processes under study. These and other results presented provide new insights for effective application of RF-based streamflow forecasting.

Introduction

Streamflow forecasting is essential for flood warnings and water resource management, but it is notoriously difficult due to the complex, non-linear interactions between precipitation, temperature, and landscape. Traditionally, hydrologists rely on "physical models" that simulate these processes but require intensive manual parameterization. Machine Learning (ML) offers a data-driven alternative, yet most ML research in hydrology has focused on rainfall-dominated regions. This paper addresses the gap by testing how a popular ML algorithm—Random Forest—performs across different hydrologic regimes, specifically comparing watersheds driven by rain, snowmelt, or a "transient" mix of both.

Methods

The researchers employed Random Forest (RF) regression to predict daily streamflow with a 1-day lead time.

  • Model Setup: Eight predictors were used, including previous day’s flow ($Q_t$), precipitation (1-day and 3-day sums), snow water equivalent (SWE), temperature (min/max), and a "pentad" (time-of-year) index.
  • Tuning: The model was optimized using an exhaustive search for the mtry parameter (number of predictors considered at each split) and a fixed ntree of 2000.
  • Benchmarks: RF was compared against a Naïve model (persistence: tomorrow's flow = today's flow) and Multiple Linear Regression (MLR).
  • Evaluation: Performance was measured using the Kling–Gupta efficiency (KGE), $R^2$, and error metrics (MAE, RMSE). They also used Probability of Detection (POD) to see how well the model caught extreme high-flow events.

Datasets and Study Area

The study utilized a decade of data (2009–2018) from 86 watersheds in the Pacific Northwest (HUC-17 region).

Study area in the Pacific Northwest (HUC-17 region).

Data sources included:

  • Streamflow: USGS National Water Information System (NWIS).
  • Meteorology: PRISM (4 km gridded precipitation and temperature data).
  • Snow Data: SNOTEL (Snow Telemetry) stations for SWE and temperature.
  • Catchment Traits: GAGES-II dataset for physical attributes like slope, soil sandiness, and forest cover.

Key Results and Insights

The study categorized watersheds into three regimes based on the timing of flow: rainfall-dominated, transient, and snowmelt-dominated.

Representative hydrographs for (b) rainfall-dominated, (c) transient, and (d) snowmelt-dominated watersheds.

Performance Across Regimes

  • Overall Accuracy: RF performed well across the board, with median KGE scores of 0.84 (rainfall), 0.87 (transient), and 0.94 (snowmelt).
  • Persistence vs. Skill: In snowmelt-dominated basins, streamflow is highly "persistent" (today looks a lot like tomorrow). This leads to very high absolute scores for all models. However, the largest improvements over benchmarks occurred in rainfall-dominated basins, which are "flashier" and more non-linear.
  • Predictor Dominance: Previous day’s flow ($Q_t$) was the most critical predictor in all cases. In snowmelt basins, snow indices (SWE and melt) were significantly more important than raw precipitation data.

Physical and Technical Constraints

  • Geography Matters: Performance deteriorated in watersheds with steeper slopes and sandier soils. In these environments, water moves too quickly for a 1-day lag model to capture the response accurately.
  • Variable Importance Conflict: The authors highlight a technical trap: two common ways to measure "variable importance" in RF (Mean Decrease in Accuracy vs. Mean Decrease in Node Impurity) yielded different rankings. They argue that researchers must interpret these through a lens of physical hydrology.

Limitations and Future Work

While RF shows promise, several areas for improvement remain:

  • Lead Time: The study only looked at a 1-day lead time. As lead time increases, the model's reliance on weather forecasts will grow.
  • Uncertainty: RF provides a point estimate but doesn't inherently quantify uncertainty. Future work should integrate probabilistic intervals.
  • Climate Change: As snowmelt-driven basins transition toward rainfall-driven regimes due to warming, the model's performance characteristics may shift, requiring adaptive retraining.

Summary updated

Abstract

In the past decades, data-driven machine-learning (ML) models have emerged as promising tools for short-term streamflow forecasting. Among other qualities, the popularity of ML models for such applications is due to their relative ease in implementation, less strict distributional assumption, and competitive computational and predictive performance. Despite the encouraging results, most applications of ML for streamflow forecasting have been limited to watersheds in which rainfall is the major source of runoff. In this study, we evaluate the potential of random forests (RFs), a popular ML method, to make streamflow forecasts at 1 d of lead time at 86 watersheds in the Pacific Northwest. These watersheds cover diverse climatic conditions and physiographic settings and exhibit varied contributions of rainfall and snowmelt to their streamflow. Watersheds are classified into three hydrologic regimes based on the timing of center-of-annual flow volume: rainfall-dominated, transient, and snowmelt-dominated. RF performance is benchmarked against naïve and multiple linear regression (MLR) models and evaluated using four criteria: coefficient of determination, root mean squared error, mean absolute error, and Kling–Gupta efficiency (KGE). Model evaluation scores suggest that the RF performs better in snowmelt-driven watersheds compared to rainfall-driven watersheds. The largest improvements in forecasts compared to benchmark models are found among rainfall-driven watersheds. RF performance deteriorates with increases in catchment slope and soil sandiness. We note disagreement between two popular measures of RF variable importance and recommend jointly considering these measures with the physical processes under study. These and other results presented provide new insights for effective application of RF-based streamflow forecasting.
BibTeX
@inproceedings{pham2021evaluation,
  author = {Leo Triet Pham and Lifeng Luo and Andrew O. Finley},
  title = {{Evaluation of random forests for short-term daily streamflow forecasting in rainfall- and snowmelt-driven watersheds}},
  booktitle = {Hydrology},
  year = {2021},
  doi = {10.5194/hess-25-2997-2021},
}
Figure