2 results
Filters
Settings
Now showing1 - 2 of 2
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: with files , Beyond prediction accuracy: Causality-driven machine learning models for seasonal flow forecasting in Canada(Elsevier, 2026-08-21); ; Toth, ElenaStudy region: Nelson–Churchill River Basin, Canada, spanning mountain, prairie, and non-prairie subbasins from the Rocky Mountains to Hudson Bay. Study focus: This study examines the role of causally informed input selection in machine learning (ML)-based seasonal streamflow forecasting. Two approaches were used to identify predictors: partial mutual information (PMI) and convergent cross-mapping (CCM). The selected inputs were used to develop two widely used ML models, Random Forest (RF) and Long Short-Term Memory (LSTM), for forecasting monthly streamflow at lead times of one to three months. Models’ performance was evaluated across multiple natural subbasin groups using Nash–Sutcliffe Efficiency (NSE) and Kling–Gupta Efficiency (KGE). New hydrological insights: Although the two approaches produced different predictor sets, ML performance was comparable across subbasin groups based on NSE and KGE, suggesting that multiple equally likely ML model configurations potentially exist for the same forecasting problem. However, the causality-based inputs demonstrated greater physical relevance compared to the information-theoretic set. This conclusion was consistent with the results of ML-based explainability analyses based on RF-based feature-importance and SHapley Additive exPlanations (SHAP) analyses. The results highlight that the first critical step in ML modeling is to identify sensible, justifiable, and physically meaningful sets of input variables to inform ML modeling. While this study offers insights into causality-based ML, a continued quest for physically meaningful ML models is needed.Publicationtype: Journal Article6 6 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Spatial variability of precipitation extremes over Italy using a fine-resolution gridded product(Elsevier, 2021-09-16); ; ; Study region: Italy. Study focus: Knowing magnitude and frequency of extreme precipitation is necessary to reduce their impact on vulnerable areas. Here we investigate the performance of the Generalized Extreme Value (GEV) distribution, using a fine-resolution satellite-based gridded product, to analyze 13,247 daily rainfall annual maxima samples. A non-extreme value distribution with a power-type behavior, that is, the Burr Type XII (BrXII), is also evaluated and used to test the reliability of the GEV in describing extreme rainfall. New hydrological insights for the region: (1) in 44.9 % of the analyzed samples the GEV predicts an upper rainfall limit; we deem this is an artifact due to sample variations; (2) we suggest the GEV<sup>+</sup> distribution, that is, the GEV with shape parameters restricted only to positive values as a more consistent model complying with the nature of extreme precipitation; (3) GEV, GEV<sup>+</sup>, and BrXII performed equally well in describing the observed annual precipitation, yet all distributions underestimate the observed sample maximum; (4) the BrXII, for large return periods, predicts larger rainfall amounts compared to GEV indicating that GEV estimates could underestimate the risk of extremes; and (5) the correlation between the predicted rainfall and the elevation is investigated. Based on the results of this study, we suggest instead of using the classical GEV to use the GEV<sup>+</sup> and non-extreme value distributions such as the BrXII to describe precipitation extremes.Publicationtype: Journal Article2