Probabilistic long‐term hydrological drought forecast using Bayesian networks and drought propagation

Abstract

Effective drought mitigation plans that can handle severe drought conditions require reliable drought forecasts. A probabilistic hydrological drought forecasting method was developed using Bayesian networks that incorporate dynamic model predictions and a drought propagation relationship. The resulting model, Bayesian networks based drought forecasting with drought propagation (BNDF_DP), was designed using current and forecast lead time drought conditions of a multi‐model ensemble. Hydrological drought conditions were represented by the Palmer Hydrological Drought Index. The ranked probability score (RPS) and receiver operating characteristic (ROC) curve analysis were employed to measure forecast proficiency. The BNDF_DP model showed good performance, with an RPS 4–50% higher than a climatological model. ROC analysis indicated that the BNDF_DP offered superior forecasting skills for long‐term drought, with a 2 and 3 month lead time, compared with a model that does not consider drought propagation. The overall results indicated that the BNDF_DP model was a promising tool for probabilistic drought forecasting that can provide water managers and decision‐makers with the flexibility to respond to undesirable drought risks, prepare drought mitigation action plans and regulate policies based on future uncertainties.

Receiver Operating Characteristic Curve Analysis-Based Evaluation of GCMs Concerning Atmospheric Teleconnections

“…Mason and Graham [93] used ROC curve analysis to validate forecast skill of GCMs concerning rainfall over the eastern part of Africa. Shin et al [94] used it to evaluate the forecast performance of a hydrological drought model. De Castro Santos et al [95] classified the intensities of El Niño episodes by applying ROC technique.…”

A seasonally blended and regionally integrated drought index using Bayesian network theory

“…BNs are a type of stochastic predictive model that illustrates the conditional dependences of a set of input random variates using direct acyclic graphs (DAGs). BNs have vast application in estimation, reasoning with uncertain evidence (Peng et al ., 2010), diagnostic applications in manufacturing (Lewis and Ransing, 1997), classification (Agrahari et al ., 2018) and drought forecasting (Kim et al ., 2018; Shin et al ., 2019). A DAG represents the sequence of events in a direct ordering.…”

Machine Learning for Understanding Inland Water Quantity, Quality, and Ecology

“…Alternatively, accuracy might be most usefully assessed with respect to values that are computed from the model predictions before being used by the target audience -for example, fisheries managers may be most interested in values computed from daily predictions, such as growing degree days, annual nutrient loads, or the probability of an algal toxin exceedance in a given week. Accuracy in predicting threshold exceedances may be assessed with Receiver Operating Characteristic (ROC) curves and/or Heidke Skill Scores (HSS), with the ROC curves providing a more nuanced view of event prediction by evaluating model sensitivity (false positive rate) against model specificity (false negative rate) for a range of thresholds (Shin et al, 2020). Continuous Ranked Probability Scores (CRPS) can also be used to evaluate probabilistic predictions and are analogous to mean absolute error (Gneiting et al, 2005;Thomas et al, 2020).…”