To map water levels related to large floods, we propose using geographical information systems to manage the vast amount of information extracted from aerial photographs. Our approach is divided into three parts : (1) segmentation of the flood plain into sectors with acceptable mean water depth, based on geographical limite and size criteria ; (2) determination of minimum and maximum depths for each sector, based on the potentiel emergence of naturel objects (vegetation, dikes, etc.) ; (3) improvement of water depth assessment, using all observed hydraulic connections between sectors (cracked dikes, flows, etc.). This advance is made possible by solving a constraint system linking the different estimated water levels. This method was applied the November 1994 flood involving the Hérault River (France), using aerial photographs taken 10 h after the flood peak. Results indicate an accurate estimation of water levels, and illustrate the lateral dynamic behaviour of the flood (1 m of difference in levels between centre and edges of plain), and the impact of dikes on water levels. These results were obtained without using either general hydraulic equations or local head loss coefficients. The great quantity of fuzzy information made up for its poor quality.