Covariates used in the stream temperature model.

Supporting sustainable lotic ecosystems and thermal habitats requires estimates of stream temperature that are high in scope and resolution across space and time. We combined and enhanced elements of existing stream temperature models to produce a new statistical model to address this need. Contrasting with previous models that estimated coarser metrics such as monthly or seasonal stream temperature or focused on individual watersheds, we modeled daily stream temperature across the entire calendar year for a broad geographic region. This model reflects mechanistic processes using publicly available climate and landscape covariates in a Generalized Additive Model framework. We allowed covariates to interact while accounting for nonlinear relationships between temporal and spatial covariates to better capture seasonal patterns. To represent variation in sensitivity to climate, we used a moving average of antecedent air temperatures over a variable duration linked to area-standardized streamflow. The moving average window size was longer for reaches having snow-dominated hydrology, especially at higher flows, whereas window size was relatively constant and low for reaches having rain-dominated hydrology. Our model’s ability to capture the temporally-variable impact of snowmelt improved its capacity to predict stream temperature across diverse geography for multiple years. We fit the model to stream temperatures from 1993–2013 and predicted daily stream temperatures for ~261,200 free-flowing stream reaches across the Pacific Northwest USA from 1990–2021. Our daily model fit well (RMSE = 1.76; MAE = 1.32°C). Cross-validation suggested that the model produced useful predictions at unsampled locations across diverse landscapes and climate conditions. These stream temperature predictions will be useful to natural resource practitioners for effective conservation planning in lotic ecosystems and for managing species such as Pacific salmon. Our approach is straightforward and can be adapted to new spatial regions, time periods, or scenarios such as the anticipated decline in snowmelt with climate change.

维持可持续流水生态系统与热栖息地，需要具备高时空范围与分辨率的河流水温估算数据。为此，我们整合并优化了现有河流水温模型的相关模块，构建了全新的统计模型以满足该需求。与此前仅能估算月/季尺度河流水温等粗粒度指标，或仅聚焦单一流域的模型不同，我们针对大范围地理区域构建了覆盖全年的逐日河流水温模型。本模型基于广义可加模型（Generalized Additive Model）框架，利用公开可得的气候与景观协变量，刻画了相关机理过程；模型允许协变量间存在交互作用，同时考量了时空协变量间的非线性关联，以更精准地捕捉季节变化模式。为表征对气候敏感性的时空差异，我们采用了前期气温的滑动平均方法，滑动窗口时长随面积标准化径流量动态调整：对于融雪主导水文的河段，尤其是高流量场景下，滑动窗口时长更长；而降雨主导水文的河段，滑动窗口时长则相对恒定且较短。本模型能够捕捉融雪的时变影响，因此提升了其在多样地理区域内的多年水温预测能力。我们基于1993至2013年的河流水温数据拟合模型，并对1990至2021年美国太平洋西北地区约261200个天然通流河段的逐日水温进行了预测。本逐日模型拟合效果优异（均方根误差RMSE=1.76，平均绝对误差MAE=1.32℃）。交叉验证结果表明，本模型可在多样景观与气候条件下的未采样点位生成可靠预测结果。该河流水温预测数据可为自然资源从业者开展流水生态系统高效保护规划，以及太平洋鲑鱼等物种的管理工作提供有力支撑。本方法简洁易用，可推广应用至新的空间区域、时间周期，或是气候变化下融雪量预期下降等未来情景。