Measured dissolved oxygen in US rivers fell less than oxygen solubility in 1981–2025 samples and sensors, while the Zhi et al. (2023) reconstruction follows solubility

Question: do measured US river oxygen data show the deoxygenation that the deep-learning reconstruction of Zhi et al. (2023, Nature Climate Change) reports? Approach: we calculated site trends (Theil-Sen, month anomalies) in two measured data sets. The first is the grab samples that Zhi et al. published as US training data (311 rivers, 1981-2019). The second is USGS continuous-sensor daily means (131-275 sites, 2000-2025). From measured water temperature we calculated air-saturation oxygen (solubility only), and we split the DO trend into a solubility part and a percent-saturation part. Intervals are 95% spatial block bootstrap intervals over 3x3 degree cells. Results: in the training samples, measured DO did not decrease (+0.024 mg/L per decade, -0.013 to +0.058; 47% of sites negative). The model output on the same dates decreased (-0.026, -0.036 to -0.018; 68% negative), and the difference is +0.050 (+0.016 to +0.083). The model and the samples agree on the solubility loss (-0.032 and -0.028 mg/L per decade). But measured percent saturation increased by 0.47 (0.11 to 0.80) % per decade, while the model shows 0.04. Model DO trends are 0.84 times the solubility trends. The sensors do not have the daytime sampling bias, and they show the same pattern. DO fell 0.098 (0.031 to 0.169) mg/L per decade less than solubility predicts in 2000-2025 and 0.073 (0.027 to 0.121) less in 2010-2025, and percent saturation increased in both windows. Summer DO and daily-minimum DO did not decrease. Conclusion: warming reduced oxygen solubility, but a simultaneous saturation increase kept measured mean DO from decreasing significantly. The reconstructed widespread US deoxygenation is mainly a solubility signal that the measured data do not show. Limits: sensor sites are not random, and the sensor windows are short. Sensor types changed in about 2005-2015. Grab-sample timing can change. We do not identify the cause of the saturation increase, and we did not test the Central European rivers or the projections.

dissolved oxygen · river deoxygenation · oxygen solubility · percent saturation · water temperature · USGS NWIS · continuous sensors · deep learning reconstruction · model validation · climate warming · trend analysis · freshwater