Abstract: Accurate streamflow prediction holds pivotal implications for water resources management and flood early warning. Nonetheless, the highly nonlinear nature of streamflow processes poses significant challenges to conventional models, which also demonstrate insufficient integration of spatiotemporal features and deficiency of interpretability. In this study, 24 types of multi-source heterogeneous data underwent systematic investigation, comprising remote sensing and meteorological data. Aside from that, the impacts of human activities and climate change were comprehensively considered to construct a high-precision and interpretable Transformer-KAN-LEC （TKL） deep learning ensemble streamflow prediction model. Taking daily streamflow prediction at 11 stations in the Jialing River Basin as a case study, the experimental findings illustrate that: the Nash-Sutcliffe efficiency coefficient （E_NS） of the TKL model is all greater than 0.95, the root mean square error （E_RMS） is reduced by 40%—80% in contrast to traditional models, and both the interval prediction and extreme event prediction performances are superior to traditional models. Interpretability analysis reveals that upstream streamflow and cumulative precipitation effects are the dominant influencing factors. The "data-model-interpretation" systematic framework recommended in this paper can offer adequate and continuous support for water resources management and precise flood early warning in large basins.