Machine Learning Models Predicting Solute Concentrations of One Stream Fail to Predict Solute Concentrations in Other Streams

Continuous sensor data is collected at 10 sites throughout New Hampshire. At each site data is collected every 15 minutes by the datalogger from a HOBO Stage logger (or site is paired with USGS site), Satlantic SUNA and YSI EXO2. Data is collected and transmitted to UNH by cell telemetry once a day where it is stored on the NH EPSCoR data server. The data that is collected from the SUNA are Nitrate in mg N/L and is corrected by grab sample NO3 analyzed by IC. Data that is collected by the YSI are Stream Temperature, pH, Specific Conductance, Dissolved Oxygen, and fDOM in QSU. The fDOM is corrected by temperature, turbidity (not included), and absorbance.

Solute concentrations from primarily weekly grab samples were used to train RF models, functioning as dependent variables in all models. Laboratory measured solutes included calcium (Ca2+), chloride (Cl−), dissolved organic carbon (DOC), potassium (K+), magnesium (Mg2+), sodium (Na+), nitrate (NO3−), and sulfate (SO42−). For the Hubbard Brook sites (W3 and W9) sample analysis occurred at the US Forest Service (Hubbard Brook Watershed Ecosystem Record (HBWatER), 2023). In brief, Cl−, NO3−, and SO42− at all sites and major cations in all but the Hubbard Brook sites were measured via ion chromatography (Pfaff & Hautman, 1999). For the Hubbard Brook sites, Ca2+, K+, Mg2+, and Na+ were measured by inductively coupled plasma atomic emission spectrometry (Pfaff & Hautman, 1999). Dissolved organic carbon and total dissolved nitrogen were measured by high temperature catalytic oxidation (Merriam et al., 1996; Potter & Wimsatt, 2005). For sites BDC, BEF, DCF, LMP72, MCQ, SBM, TPB, and WHB, water chemistry analyses were performed at the Water Quality Analysis Laboratory at the University of New Hampshire (Wymore et al., 2021). Anions (NO3−, Cl−, SO42−) and cations (Na+, Ca2+, Mg2+, K+) were analyzed using ion chromatography with suppressed conductivity detection (Dionex 1000 ICS with autosampler; Sunnyvale, CA). Units of NO3− and SO42− refer to the mass per unit volume of N and S, respectively. DOC (NDIR detection) and TDN (chemiluminescent detection) were measured by high-temperature catalytic oxidation with a Shimadzu TOC-LCSH with a TNM-1 nitrogen unit. Dissolved organic nitrogen (DON) was estimated via subtraction: DON = Total Dissolved Nitrogen [TDN] – Dissolved Inorganic Nitrogen [DIN]) where DIN equals the sum of NH4+ and NO3-.

Methods and findings are described in the associated manuscript.

本数据集采集了覆盖新罕布什尔州全境10个监测点位的连续传感器数据。各点位每15分钟由数据记录仪（datalogger）从HOBO水位记录仪（HOBO Stage logger，部分点位配对美国地质调查局（USGS）站点）、Satlantic SUNA传感器及YSI EXO2分析仪采集数据。数据通过蜂窝遥测技术每日采集并传输至新罕布什尔大学（UNH），并存储于新罕布什尔州EPSCoR（NH EPSCoR）数据服务器中。从Satlantic SUNA采集的数据为以mg N/L为单位的硝酸盐浓度，该数据通过离子色谱（ion chromatography，IC）分析的即时水样NO3进行校正。YSI EXO2采集的数据则包括河流水温、pH值、比电导率、溶解氧及以QSU为单位的荧光溶解有机物（fluorescent dissolved organic matter，fDOM）浓度。荧光溶解有机物（fDOM）数据通过温度、浊度（未纳入本数据集）及吸光度进行校正。

本数据集主要以每周采集的即时水样中的溶质浓度作为因变量，用于训练随机森林（Random Forest，RF）模型。实验室测定的溶质包括钙离子（Ca²+）、氯离子（Cl⁻）、溶解性有机碳（dissolved organic carbon，DOC）、钾离子（K+）、镁离子（Mg²+）、钠离子（Na+）、硝酸根（NO3⁻）及硫酸根（SO4²⁻）。对于哈伯德布鲁克监测点位（W3与W9），水样分析由美国林务局（US Forest Service）完成，相关记录详见《哈伯德布鲁克流域生态系统记录（HBWatER），2023》。简言之，所有点位的氯离子、硝酸根及硫酸根，以及除哈伯德布鲁克点位外所有监测点的主要阳离子均通过离子色谱（IC）测定（Pfaff & Hautman，1999）。而哈伯德布鲁克点位的钙离子、钾离子、镁离子及钠离子则通过电感耦合等离子体原子发射光谱法测定（Pfaff & Hautman，1999）。溶解性有机碳与总溶解氮则通过高温催化氧化法测定（Merriam等，1996；Potter & Wimsatt，2005）。对于BDC、BEF、DCF、LMP72、MCQ、SBM、TPB及WHB点位，水化学分析由新罕布什尔大学水质分析实验室完成（Wymore等，2021）。阴离子（硝酸根、氯离子、硫酸根）与阳离子（钠离子、钙离子、镁离子、钾离子）采用带抑制型电导检测的离子色谱法进行分析，仪器配置为带自动进样器的Dionex 1000 ICS（美国加利福尼亚州森尼韦尔市）。硝酸根与硫酸根的浓度单位分别对应氮与硫的单位质量体积浓度。溶解性有机碳（采用非分散红外检测（non-dispersive infrared，NDIR））与总溶解氮（采用化学发光检测（chemiluminescent detection））通过搭载TNM-1氮单元的岛津（Shimadzu）TOC-LCSH总有机碳分析仪进行高温催化氧化测定。溶解有机氮（dissolved organic nitrogen，DON）通过差值法估算：DON = 总溶解氮（total dissolved nitrogen，TDN） – 溶解无机氮（dissolved inorganic nitrogen，DIN），其中溶解无机氮为铵根（NH4+）与硝酸根的浓度之和。

本数据集的研究方法与结果详见相关学术论文。