An Automatic System for Flooded Area Detection: A Machine Learning Approach

Floods are some of the most destructive and catastrophic natural events, as well as the most recurrent. This works objective is to explore the capabilities of a powerful tool as Machine Learning in order to build an automatic system for the detection of flooded area analyzing geospatial data collected from satellitar missions. To achieve this, different systems (based on several Machine Learning algorithms) are built and tested, in order to make a comparison of their performances on the the data set. After the test of state of the art techniques as Convolutional Neural Networks, other white box models are experimented, in order to collect informations about the data set, and on the most important features towards the detection of flooded area. Moreover, different image processing tools were applied in pre-processing and post-processing phase and evaluated in relation to each methods performance variations. This Thesis contains the report of all the experiments performed, a deep analysis of the results collected, and the consequential considerations and evidences they have pointed out.