The present work aims to assess how the flood hazard maps obtained with hydraulic approaches can be retrieved with machine leaning techniques. The presented case study is the Argens basin in SouthEastern France. The reference flood hazard maps cover a 1163.1 km length of rivers, and were obtained through the simulation of water depth for a 1000-year return period peak discharges, using the FLOODOS 2d hydraulic model at a 5-m resolution. From these hazard maps, a flood inventory combining an equal number of flood and non-flood points was generated, and was divided through random sampling into training and validation datasets with 70:30 ratio for the application of machine learning methods. The explanatory variables selected for learning include several geo-environmental factors (GeFs): 5m resolution Digital Terrain Model (DTM) and Height Above Nearest Drainage (HAND) rasters; average annual rainfall on a 1x1km grid; and geology map; 1000-year river discharge dataset used for the reference hydraulic modeling. The multi-collinearity analysis was used to opt out the GeFs having a collinearity issue. Advanced artificial neural network, random forest, and extreme gradient boosting models were trained on the data sets. To validate these models, we have considered the area under the receiver operating characteristic (AUROC), a widely used validation technique, and also another metric focusing on the accuracy the estimated flood extent, i.e., the Critical Success Index. The results illustrate the performances that can be achieved with machine learning for flood mapping in areas with complex geo-topographic features such as the Argens basin.