Characterization of Time-Varying Regimes in Remote Sensing Time Series: Application to the Forecasting of Satellite-Derived Suspended Matter Concentrations

The spatial and temporal coverage of satellites provides data that are particularly well suited for the analysis and characterization of space–time-varying geophysical relationships. The latent-class models aim here to identify time-varying regimes within a dataset. This is of particular interest for geophysical processes driven by the seasonal variability. As a case example, we study the daily concentration of mineral suspended particulate matters estimated from satellite-derived datasets, in coastal waters adjacent to the French Gironde river mouth. We forecast this high-resolution dataset using environmental data (wave height, wind strength and direction, tides, and river outflow) and four latent-regime models: homogeneous and nonhomogeneous Markov-switching models, with and without an autoregressive term (i.e., the mineral suspended matter concentration observed the day before). Using a validation dataset, significant improvements are observed with the multiregime models compared to a classical multiregression and a state-of-the-art nonlinear model [support vector regression (SVR) model]. The best results are reported for a mixture of three regimes for the autoregressive model using nonhomogeneous transitions. With the autoregressive models, we obtain at day+1 for the mixture model forecasting performances of 93% of the explained variance, compared to 83% for a standard linear model and 85% using an SVR. These improvements are more important for the nonautoregressive models. These results stress the potential of the identification of geophysical regimes to improve the forecasting. We also show that nonhomogeneous transition probabilities and estimated autoregressive terms improve forecasting performances when observation data is lacking for short-time period of 1–15 days.