Question: does the month-to-month sensitivity of stream temperature to air temperature and discharge predict the decadal change? Approach: at USGS NWIS sites with long daily water temperature records, we fitted the sensitivity on monthly anomalies in 2000-2012 only and predicted the change in mean water temperature to 2013-2025. Air temperature came from GHCN_CAMS (primary), nClimGrid and Berkeley Earth. Intervals come from a 5-degree spatial block bootstrap. Result: for July-August at 356 sites, the ratio of mean observed to mean predicted change was 0.86 (0.40 to 1.20) with GHCN_CAMS, 1.05 with nClimGrid and 1.11 with Berkeley Earth, and the predictions reduced site-level squared error by 30% relative to no change. Annual means warmed more than predicted (ratio 1.26 to 1.52 over the three products). At 281 year-round sites, the predicted monthly pattern of change agreed with the observed pattern (correlation 0.92 to 0.96), but November-March water warmed 0.11 C (0.04 to 0.17) more than predicted. This excess was mainly east of 100 W (cold-minus-warm residual +0.14 to +0.23 C; West near 0). A placebo (2000-2006 vs 2007-2012, sensitivity from 2013-2025) gave no all-site or eastern contrast, but it gave a western contrast of +0.14 C, so regional residuals near 0.1 C can occur in this design. The excess was not larger at ice-affected or high-baseflow sites and did not increase with local air warming, so a simple groundwater-lag explanation does not fit. Limitations: non-random, partly regulated sites; 13-year periods; gridded air temperature; linear monthly model; the cause of the cold-season excess is not identified.
Filters: anzsrc-for-2020: 370704 ×
Austin and Colman (2007) reported that Lake Superior July-September surface water warmed about 1.1 C per decade in 1979-2006, about two times faster than regional air, and gave an ice-albedo explanation. We did an out-of-sample test with the same three NDBC buoys, six buoys on Lakes Michigan, Huron, Erie and Ontario, GLERL daily ice cover and GHCN-CAMS basin air temperature. All models were fitted on 1979-2006 and tested on 2007-2025 (18 Superior seasons). We reproduced the 1979-2006 trend (+1.16 C per decade, 95% CI +0.24 to +2.07). In 2007-2025 the trend was +0.43 (-0.75 to +1.60), and the extrapolated trend line overpredicted the test mean by 1.58 C (RMSE 2.84 C, against 2.58 C for the training mean). A linear model with December-March, April-June and July-September air temperature (coefficient sum 2.27 C per C) predicted the test years with RMSE 1.20 C and bias +0.01 C, 1.63 C (95% CI 0.99 to 2.27) better than the trend line. With air temperature in the model, Jan-Apr ice cover (r = -0.86 with winter air) had no effect (-0.001 C per percentage point, 95% CI -0.033 to +0.032) and did not improve test or cross-validated error. The air model also explained the 1979-2025 trend (residual year term +0.06 C per decade). Michigan, Huron and Erie gave the same forecast result; Michigan and Huron showed a small in-sample ice effect (about -0.2 C per SD) that cross-validation did not clearly confirm. Limitations: short test record, point buoys near the surface, observed (not forecast) air in the test, and ice and winter air change together, so the test cannot separate an ice-albedo path from other winter heat-loss paths.
Question: does the 2024 LSU model, which makes Gulf of Mexico hypoxic area per May nitrate load a curved function of May air temperature at New Orleans, predict areas outside its 2000-2023 fit years? Approach: we used USGS daily WRTDS-K loads for the Mississippi and Atchafalaya rivers (ver. 3.0, 2026), LUMCON areas for 1985-2026, GHCN-Daily temperature, HURDAT2 storms and NDBC winds. We reproduced the model, froze it, applied it to 1985-1999 and 2024-2025, did month-placebo tests and made rolling one-year-ahead forecasts. Results: the first operational forecast (2024) was 5,107 km2 (95% interval 4,687-5,569). The measured area was 17,360 km2, and a load-only model gave 14,113 km2. A storm cannot explain an area that is larger than the forecast. In the 1985-1999 back-cast (12 years), the temperature model increased the mean absolute error from 4,493 to 6,606 km2 (difference CI +826 to +3,432). In the fit years, February and November temperatures fit the area-per-load ratio as well as May. Over 30 storm-free years, a 1 C warmer May gave 11% less area per load (p = 0.046), and a warmer June gave 23% less. River discharge did not remove this link. In rolling forecasts for 18 storm-free years (2000-2025), the curved May model decreased the error by 451 km2 (CI 50-851). May was the only one of 13 predictors with this result, and the authors selected it with the same data. A monotonic May term gave no clear gain. 15 of 18 published LSU forecasts from 2007 to 2026 were larger than the measured area. Limitations: one area value per year, three years after the fit with storms in two, and the exact model equation is not published.
Question: does a linear trend fitted to past annual flood peaks forecast later peaks better than a stationary model? We used USGS annual peak discharges at 1,187 gauges with no regulation, diversion or urbanization code, at least 55 peaks in water years 1941-2005 and at least 16 peaks in 2006-2025 (264 are CAMELS reference basins). We fitted six models to ln(peak) in 1941-2005 and scored 2006-2025 forecasts with the continuous ranked probability score (CRPS). The plan was written before scoring. The at-site linear trend was worse than the stationary model: skill -2.9% (95% interval from test-year resampling -4.6% to -1.3%); -4.6% at CAMELS gauges. Its point estimate was negative in all eight regional groups. A trend used only when Mann-Kendall p<0.05 gave -1.1%, a regional median trend -0.3%, the 1976-2005 mean +0.8% and half of the trend +0.9%. Full trends also lost 5-10% at three other forecast origins (post hoc). Across gauges, the 2006-2025 anomaly increased by 0.38 (0.30-0.46) per unit of trend-predicted shift, below the 0.56-0.63 that estimation noise alone would give; so trends continued only in part. A post hoc scan found the best share of the trend at 30-40%, with about 1.1% gain. Stationary 10-year and 50-year floods were exceeded by 10.1% and 2.5% of 2006-2025 peaks. At CAMELS gauges, test-period anomalies followed annual precipitation anomalies, but precipitation explained only 7% of the variance of old trends. Limitations: lognormal predictive distributions, linear mean trends only, coding-based screening of human effects, 20 shared test years, and post hoc secondary analyses.