The present paper shows that a wide class of complex transfer functions encountered in geophysics can be efficiently modeled using neural networks. Neural networks can approximate numerical and nonnumerical transfer functions. They provide an optimum basis of nonlinear functions allowing a uniform approximation of any continuous function. Neural networks can also realize classification tasks. It is shown that the classifier mode is related to Bayes discriminant functions, which give the minimum error risk classification. This mode is useful for extracting information from an unknown process. These properties are applied to the ERS 1 simulated scatterometer data. Compared to other methods, neural network solutions are the most skillful.

This paper presents a new method for segmenting multispectral satellite images. The proposed method is unsupervised and consists of two steps. During the rst step the pixels of a learning set are summarized by a set of codebook vectors using a Probabilistic Self-Organizing Map (PSOM, 9]) In a second step the codebook vectors of the map are clustered using Agglomerative Hierarchical Clustering (AHC,7]). Each pixel takes the label of its nearest codebook vector. A practical application to Meteosat images illustrates the relevance of our approach.

No abstract

Abstract. Variational data assimilation consists in estimating control parameters of a numerical model in order to minimize the misfit between the forecast values and some actual observations. The gradient based minimization methods require the multiplication of the transpose jacobian matrix (adjoint model), which is of huge dimension, with the derivative vector of the cost function at the observation points. We present a method based on a modular graph concept and two algorithms to avoid these expensive multiplications. The first step of the method is a propagation algorithm on the graph that allows computing the output of the numerical model and its linear tangent, the second is a backpropagation on the graph that allows the computation of the adjoint model. The YAO software implements these two steps using appropriate algorithms. We present a brief description of YAO functionalities.

[1] The global ocean biogeochemical models that are used in order to assess the ocean role in the global carbon cycle and estimate the impact of the climate change on marine ecosystems are getting more and more sophisticated. They now often account for several phytoplankton functional types that play particular roles in marine food webs and the ocean carbon cycle. These phytoplankton functional types have specific physiological characteristics, which are usually poorly known and therefore add uncertainties to model results. Indeed, this evolution in model complexity is not accompanied by a similar increase in the number and diversity of in situ data sets necessary for model calibration and evaluation. Thus, it is of primary importance to develop new methods to improve model performance using existing biogeochemical data sets, despite their current limitations. In this paper, we have optimized 45 physiological parameters of the PISCES global model, using a variational optimal control method. In order to bypass a global 3-D ocean variational assimilation, which would require enormous computation and memory storage, we have simplified the estimation procedure by assimilating monthly climatological in situ observations at five contrasted oceanographic stations of the JGOFS program in a 1-D version of the PISCES model. We began by estimating the weight matrix in the cost function by using heuristic considerations. Then we used this matrix to estimate the 45 parameters of the 1-D version of the PISCES model by assimilating the different monthly profiles (observed profiles at the five stations) in the same variational procedure on a time window of 1 year. This set of optimized parameters was then used in the standard 3-D global PISCES version to perform a 500 year global simulation. The results of both the standard and the optimized versions of the model were compared to satellite-derived chlorophyll-a images, which are an independent and global data set, showing that our approach leads to significant improvements in simulated surface chlorophyll-a in most of the regions of the world ocean. Besides demonstrating that we have improved the accuracy of the PISCES model, this study proposes a sound methodology that could be used to efficiently account for in situ data in biogeochemical ocean models.