When collections of functional data are too large to be exhaustively observed, survey sampling techniques provide an effective way to estimate global quantities such as the population mean function, without being obligated to store all the data. In this thesis, we propose a Horvitz–Thompson estimator of the mean trajectory, and with additional assumptions on the sampling design, we state a functional Central Limit Theorem and deduce asymptotic confidence bands. For a fixed sample size, we show that stratified sampling can greatly improve the estimation compared to simple random sampling. In addition, we extend Neyman’s rule of optimal allocation to the functional context. Taking into account auxiliary information has been developed with model-assisted estimators and weighted estimators with unequal probability sampling proportional to size. The case of noisy curves is also studied with the help local polynomial smoothers. To select the bandwidth, we propose a cross-validation criterion that takes into account the sampling weights. The consistency properties of our estimators are established, as well as asymptotic normality of the estimators of the mean curve. Two methods to build confidence bands are proposed. The first uses the asymptotic normality of our estimators by simulating a Gaussian process given estimated the covariance function in order to estimate the law of supremum. The second uses bootstrap techniques in a finite population that does not require to estimate the covariance function.
