A nonlinear, neural-network-based extension of the principal component analysis (PCA) is applied to the water level and current fields in a shallow tidal sea at the German North Sea coast. Contrary to the linear PCA, which tends to split patterns in the data among several modes difficult to interpret, the nonlinear PCA enables to identify the nonlinear spatial patterns in the data with only a single mode. The first nonlinear principal component (PC) corresponds well with the joint probability distribution of the linear PCs and can be argued to represent a 'typical' tidal cycle in the study area.