Extended Data Fig. 2: Distribution of cross-validation results.

a, Maps of spatially cross-validated predictive accuracy of flow intermittence for streamflow gauging stations. See Supplementary Fig. 3 for the distribution of spatial cross-validation folds and details on the cross-validation procedure. The classification errors shown here are not necessarily present in the final predictions but illustrate the ability of the model to predict the flow intermittence class for each region if that region was excluded from the training set. For instance, it shows that the model would be unable to predict the presence of IRES in western France and northern Spain (inset ii, dark red dots), or in western India (inset iii) without training stations in these regions. b–e, Intermittence prediction residuals versus gauging station characteristics and environmental variables. The mean intermittence prediction residual (IPR) is the difference between the average predicted probability of flow intermittence (across three cross-validation folds and two repetitions) and the observed flow intermittence of the gauging station (1 = non-perennial, 0 = perennial). Overall, prediction errors and uncertainties decrease with an increase in the number of recorded years by gauging stations as well as the drainage area and the degree of flow intermittence (average annual number of zero-flow days and flow cessation events) of the corresponding reaches. Mapping software: ArcMap (ESRI).