Machine Learning-Assisted Colorimetric Quantification of Nitrite in Plateau Lake Waters

Abstract

Nitrite (NO2−) is a transient intermediate in aquatic nitrogen cycling, and its rapid fluctuations are difficult to capture in heterogeneous natural waters. Here, we developed a CS5-based smartphone colorimetric workflow coupled with machine learning for on-site quantification of nitrite in plateau lake water. The CS5 assay showed a nitrite-dependent visible color change and a pronounced UV−vis absorption blue shift, accompanied by near-infrared fluorescence quenching. Smartphone RGB features served as the primary quantitative readout, with a detection limit of 0.27 μM. Across five plateau lakes, RGB-derived estimates showed a strong linear association with conventional measurements (R2 = 0.85). Sediment-water microcosms from three lakes at 15, 25, and 35 °C provided time-resolved RGB and chemical data for model development. Among six models, random forest regression yielded the lowest cross-validated and held-out RMSE for nitrite prediction. The corresponding R2 values were 0.91 and 0.84, respectively. Leave-one-lake-out validation showed that RF-assisted color-chart correction improved cross-lake prediction, increasing R2 from 0.84 to 0.89 and reducing RMSE from 3.71 to 2.10 μM. Time-resolved nitrite dynamics were further interpreted alongside nitrogen-species and enzyme-activity measurements. Together, these results establish a systematically validated, CS5-based smartphone workflow for quantitative nitrite analysis across complex lake-water matrices.