The determination of soil parameters in geotechnical engineering and their variations during the construction process have long been a focal point for engineering designers. While the artificial neural network (ANN) has been employed for back analysis of soil parameters, its application to caisson sinking processes remains limited. This study focuses on the Nanjing Longtan Yangtze River Bridge project, specifically the south anchoring of an ultra-large rectangular caisson. A comprehensive analysis of the sinking process was conducted using 400 finite element method (FEM) models to obtain the structural stress and earth pressure at key locations. Multiple combinations of soil parameters were considered, resulting in a diverse set of simulation results. These results were then utilized as training samples to develop a back-propagating artificial neural network (BP ANN), which utilized the structural stress and earth pressure as input sets and the soil parameters as output sets. The BP ANN was individually trained for each stage of the sinking process. Subsequently, the trained ANN was employed to predict the soil parameters under different working conditions based on actual monitoring data from engineering projects. The obtained soil parameter variations were further analyzed, leading to the following conclusions: (1) The soil parameters estimated by the ANN exhibited strong agreement with the original values from the geological survey report, validating their reliability; (2) The surrounding soil during the caisson sinking exhibited three distinct states: a stable state prior to the arrival of the cutting edges, a strengthened state upon the arrival of the cutting edges, and a disturbed state after the passage of the cutting edges; (3) In the stable state, the soil parameters closely resembled the original values, whereas in the strengthened state, the soil strength and stiffness significantly increased, while the Poisson’s ratio decreased. In the disturbed state, the soil strength and stiffness were slightly lower than the original values. This study represents a valuable exploration of back analysis for caisson engineering. The findings provide important insights for similar engineering design and construction projects.