Space and Time Modelling of Clouds

Abstract

Satellite images provide detailed spatial and temporal information about the structure of the atmosphere: state at a time t or evolution from time t — n to time t. The conception of a model able to forecast the motion of the clouds (in other terms, a space-time model for clouds) relies on high frequency space and time measurements of relevant features of the atmosphere. In the framework of this thesis, position, shape and brightness of the clouds are extracted from METEOSAT-6’s raw visible images via a pre-processing step consisting in the setting of a static threshold on the ‘cloud optical thickness’ of each pixel in the image. The next step in the project is the creation of a model for clouds. As clouds are assumed to be groups of ‘cloud cells’, modelling a cloud comes down to modelling each of its basis ‘cloud cells’. Thus, a cloud is modelled as a Radial Basis Function network with Gaussian radial basis functions. In the last step of the project, the dynamics of the model are investigated. The advection field, responsible (among other processes) for the moving of the cloud field, can be seen as a ‘wind flow’. Forecasting the state of the motion of the clouds from time t — 1 to time ¢ consists in forecasting the next state of the advection field as well as estimating the parameters of the Gaussian radial basis functions modelling the ‘cloud cells’. To do so, a Kalman filter-like approach has been undertaken: the filter supports estimations of the past, the present and even future states without knowing the precise nature of the modelled system.