An in-situ optical-computational soft sensor for high-resolution estuarine pCO2 estimation using dual-path absorption spectroscopy and interpretable model diagnostics

Continuous monitoring of the partial pressure of carbon dioxide ( p CO 2 ) in dynamic estuarine environments is fundamentally constrained by the low spatiotemporal resolution of discrete sampling and the stability (e.g., calibration demand, frequent maintenance) of direct-contact transducers. This creates a practical need for field-deployable workflows that combine rapid in-situ optical measurements with reference-calibrated modeling, interpretable diagnostics, and uncertainty-aware evaluation. Here, we developed and evaluated an in-situ optical-computational soft sensor for the Pearl River Estuary (PRE) by coupling a dual-path ultraviolet–visible (UV–Vis) absorption spectrometer with temperature, salinity, and machine-learning regression. The dataset comprised 3,532 synchronized valid records from five seasonal cruises, with reference p CO 2 provided by a calibrated membrane-equilibration non-dispersive infrared instrument and optical absorbance collected every 36 s. This temporally held-out evaluation provides a realistic assessment of cruise-to-cruise transferability within the PRE and shows that a compact optical input stream can support rapid p CO 2 estimation when calibrated against the membrane-NDIR reference measurements.. Model-agnostic permutation importance and PCA-loading-weighted spectral diagnostics indicated that spectral PCs and hydrographic covariates jointly contributed to prediction, linking optical variability with temperature-salinity context in a process-consistent manner. At the field scale, the calibrated soft sensor reproduced the dominant PRE p CO 2 gradients and enabled high-frequency spatial visualization of estuarine heterogeneity. These capabilities support rapid screening of CO 2 -enriched zones, cruise-scale carbon mapping, adaptive sampling design, and denser observational support for estuarine carbon-cycle assessment. Overall, the framework offers a practical and maintenance-conscious route for extending calibrated p CO 2 observations beyond discrete sampling while retaining interpretable and uncertainty-aware model diagnostics.