A Small-Sample Water Quality Retrieval Method and Its Reliability Assessment Using UAV Multispectral Remote Sensing in a Small Inland Lake

This study demonstrates a workflow for assessing the reliability of small-sample water quality retrieval from full-coverage UAV multispectral surveys of a small inland lake. Because at most 20 paired samples were available, a physically constrained feature selection strategy was developed to reduce spurious spectral relationships. Nested leave-one-out cross-validation provided out-of-sample predictions, and 95% confidence intervals and end-to-end label permutation tests assessed their statistical support. Constructed features raised the average R2 by 0.455 over raw bands. Chromophoric dissolved organic matter (CDOM), total nitrogen (TN), total phosphorus (TP), and suspended solids (SS) achieved R2 of 0.817, 0.671, 0.637, and 0.549 and satisfied both reliability criteria, whereas chlorophyll-a (Chla) and turbidity (TUB) did not, despite moderate R2 values. Under extremely limited sample sizes, higher predictive performance therefore does not necessarily indicate more reliable retrieval. However, this statistical support does not extend to the whole mapped lake, because most pixels for CDOM, TP, and SS lay outside the calibration range and some mapped values were physically inadmissible. Nor does it establish temporal or external transferability, as the findings are based on a single small inland lake with limited spatiotemporal field sampling. Their applicability to other water bodies and sampling conditions requires further validation.