Temporally transferable machine learning model for total suspended matter retrieval from Sentinel-2

The empirical (regression-based) models have long been used for retrieving water quality parameters from optical imagery by training a model between image spectra and collocated in-situ data. However, a need clearly exists to examine and enhance the temporal transferability of models. The performance of a model trained in a specific period can deteriorate when applied at another time due to variations in the composition of constituents, atmospheric conditions, and sun glint. In this study, we propose a machine learning approach that trains a neural network using samples distributed in space and time, enabling the temporal robustness of the model. We explore the temporal transferability of the proposed neural network and standard band ratio models in retrieving total suspended matter (TSM) from Sentinel-2 imagery in San Francisco Bay. Multitemporal Sentinel-2 imagery and in-situ data are used to train the models. The transferability of models is then examined by estimating the TSM for imagery acquired after the training period. In addition, we assess the robustness of the models concerning the sun glint correction. The results imply that the neural network-based model is temporally transferable (R2 ≈ 0.75; RMSE ≈ 7 g/m3 for retrievals up to 70 g/m3) and is minimally impacted by the sun glint correction. Conversely, the ratio model showed relatively poor temporal robustness with high sensitivity to the glint correction.