Victorian catchments that had not recovered from the Millennium Drought by 2017 stayed in a low-runoff state through the wet years 2020-2022 and to 2025 (143 gauges, annual hidden Markov models)

Question: Peterson et al. (2021, Science) reported that 37% of Victorian catchments had not returned to their pre-drought rainfall-runoff relation by 2017, 7 years after the Millennium Drought. Did these catchments recover in the 8 later years 2018-2025, which include the wet La Nina years 2020-2022? Approach: we used 143 catchments of the Saft Victorian set in CAMELS-AUS v2.03, Bureau of Meteorology daily flow for 1950-2025 and SILO catchment rainfall. We wrote a Python port of the default hydroState annual model (Box-Cox runoff linear in rainfall, two hidden states that change the intercept and error, optional AR1, AIC selection). We fitted it with data to 2017, then tracked and scored 2018-2025 with fixed parameters, and refitted to 2025. We wrote the plan before we calculated any output after 2017. Result: with data to 2017, 46 of 143 catchments (32%; 37% with no-AR models) were in the low state, which agrees with the original study. After the refit to 2025, 34 of these 46 (74%) were in the low state in 2022 and in 2025. For 2018-2025 annual runoff, 'stay low' predicted better than 'return to normal' by a median of +1.37 log units per year (95% CI 0.87-1.72; +1.28 in 2020-2022 only). Their mean residual relative to the pre-1997 relation was -1.74 SD in 2010-2017 and -1.58 SD in 2020-2022, although 2020-2022 rain was 11% above the pre-1997 mean. Catchments with a return had less 2022 rain than those without one. A 97% data-completeness sensitivity (133 catchments) gave the same results. Limitations: not the exact original catchments or code; state labels have a 13-24% false-alarm level before 1997; rainfall-only model; calendar years; 8 out-of-sample years.

Millennium Drought · rainfall-runoff relationship · hydrological non-stationarity · drought recovery · hidden Markov model · Victoria Australia · CAMELS-AUS · out-of-sample test · replication