Detection and attribution of eco-hydrological alteration based on deep learning-driven gap-filled runoff in a large-scale catchment

Study region: The upper Dulong-Irrawaddy River Basin (Upper DIRB) in the China-Myanmar border region, with a notable lack of hydrological observations. Study focus: Climate change and human activities have markedly altered the hydrological conditions of rivers, negatively impacting river ecosystems and regional water resources management. Especially international rivers are emblematic of the ecological stresses and water security challenges exacerbated by these alterations. Thus, this study employed an advanced and credible deep learning model to bridge the data gap. Then the eco-hydrological alterations were analyzed though a continuous and extensive time series, employing the most ecologically relevant hydrological indicators (ERHIs) for attribution analyses. New hydrological insights: Of the several models deployed (BiLSTM, LSTM, BiGRU, GRU), the bidirectional long short-term memory (BiLSTM) model performed best (NSE=0.90, KGE=0.93), which effectively reconstructed the missing daily runoff for 1989–1995. Our systematic assessment highlighted a marked disturbance in the natural hydrology post-1998, with overall hydrological change degree of 73 %. We further identified seven critical ERHIs and noted significant shifts in high and low pulse durations. The attribution analysis underscored the predominant role of human activities, accounting for 70 % of the runoff reduction and dominating five of the seven ERHIs, while climate dominated the 3-day maximum and low pulse count. These findings help regional water resources management and river ecosystem conservation and provide new insights into eco-hydrological research in data-scarce regions.