Satellite data, with their spatial and temporal coverage, are particularly well suited for the analysis and characterization of space-time-varying relationships between geophysical processes. We investigate here the forecasting of a geophysical variable using both satellite observations and model outputs. As 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 multi-latent-regime models: homogeneous and non-homogeneous Markov-switching models, with and without an autoregressive term, i.e. the suspended matter concentration observed the day before. We clearly show, using a validation dataset, significant improvements with multi-regime models compared to a classical multi-regression and a state-of-the-art non-linear model (Support Vector Regression (SVR) model). The best results are reported for a mixture of 3 regimes for autoregressive model using non-homogeneous transitions. With the autoregressive models, we obtain at day+1 forecasting performances of 93% of the explained variance for the mixture model compared to 83% for a standard linear model and 85% using a SVR. These improvements are even more important for the non-autoregressive models. These results stress the potential of the identification of geophysical regimes to improve the forecasting or the inversion. We also show that for short periods of lack of observations (typically lesser than 15 days), non-homogeneous transition probabilities and estimated autoregressive term, the observation of the previous day not being available, help to enhance forecasting performances.