Satellite data are daily acquired over Black Sea and allow visualizing the surface circulation and its meso-scale structures at the spatial resolution and time frequency of the sensor. The remotely-sensed images are in fact the only continuous source of information at fine scales greater for analysing eddies, jets, filaments, mushroom-shaped vortices... A number of applications, such as oil spill monitoring for instance, also require being able to forecast the surface circulation at short temporal horizon. This nowcasting issue includes two major components: a first module is in charge of continuously estimating the meso-scale structures of the surface dynamics and a second module is in charge of forecasting the future image data at a given temporal horizon. The presentation summarizes these two modules. It also provides results obtained with NOAA-AVHRR data acquired over Black Sea. The estimation module relies on a sliding-window approach, which processes the last few images and allows to compute a new couple (estimation, forecast), when a new image is acquired. The module computes the surface dynamics on a given temporal interval with a data assimilation approach. Based on a number of images, usually four or five, a data assimilation method estimates the surface velocity field on the whole time interval corresponding to the acquisitions from a numerical temporal evolution model. The issue of missing data, due to the cloud cover or acquisition errors for instance, is taken into account thanks to the observation error and its covariance matrix, which are included in the mathematical formulation of data assimilation. The temporal evolution model considers the transport-diffusion of the surface temperature and the constancy of velocity on the trajectory of each water. The chosen approach is named 4D-Var variational data assimilation. This method iteratively computes the initial condition of the image and velocity fields from the comparison at acquisition dates of the image fields with the real image acquisitions. Their discrepancy is included in a cost function that is iteratively minimized for computing the solution. This 4D-Var method requires the development of the adjoint of the evolution model, which is used for calculating the gradient of the cost function. The forecasting module runs a simulation software and synthetizes future images based on the previously estimated dynamics and on the last acquisitions. As the simulation runs are computed on a longer duration, compared to the one performed during the estimation phase, they rely on semi-Lagrangian schemes. The accuracy of these numerical schemes is, additionally to the performance of the surface velocity estimation, a key point for the quality of the forecasts. The basic evaluation criteria are to first compare these forecasts with a persistence hypothesis and then with the real acquisitions at same dates. This method is implemented with the Image Forecast software. The code is currently under investigation for parallelizing the most costly components and allowing short-term and reliable forecasts of the sea surface temperature images on the Black Sea domain.