Evaluation of minimal local recalibration for reducing transfer error in Sentinel-2 chlorophyll a estimation across heterogeneous waterbodies

Chlorophyll a (Chl a ) is a commonly used indicator of phytoplankton biomass, trophic status, and eutrophication risk in inland waters. Sentinel-2 satellite imagery can scale up spatiotemporal chlorophyll monitoring, but transferability of retrieval models across heterogeneous waterbodies remains a challenge for operational use. This study evaluated a local bias recalibration approach for Sentinel-2-based Chl a estimation in heterogeneous inland waters. The study covered 28 lakes and reservoirs in eastern Poland and included 203 field-satellite match-ups after quality control. Six modelling approaches were evaluated: 2 linear models, 2 generalized additive models, Elastic Net, and Random Forest. Transferability was assessed using leave-one-cluster-out cross-validation. A minimal recalibration of the bias component was tested using 0–3 reference samples. Transfer prediction errors were characterised using bias-scatter analysis. Without recalibration, generalized additive models and Random Forest showed the best transfer performance, with median RMSE of 0.270–0.274 and FAC2 of 69.05–71.43%. A single calibration sample significantly reduced transfer bias. For Linear (MCI), RMSE decreased from 0.377 to 0.190 and FAC2 increased from 35.42% to 85.71% from k = 0 to k = 1. The largest improvements were observed for simple linear models, while more complex models showed comparatively smaller gains. Recalibration mainly corrected systematic bias, with minimal effect on residual scatter. With 1–3 calibration points, simple models approached the performance of more complex models. Minimal local bias recalibration can effectively support Chl a monitoring in heterogeneous inland waters, particularly in systems that already rely on periodic in situ sampling.