Predicting flood magnitude and location helps decision-makers to better prepare for flood events. To increase the speed and availability of data during flooding, this study presents a vision-based framework for measuring water levels and detecting floods. The deep learning models use time-lapse images captured by surveillance cameras to detect water extent using semantic segmentation and to transform them into water level values with the help of lidar data.

The rise of new technologies such as drones (unmanned aerial systems – UASs) has allowed widespread use of image velocimetry techniques in place of more traditional, usually slower, methods during hydrometric campaigns. In order to minimize the velocity estimation errors, one must stabilise the acquired videos. In this research, we compare the performance of different UAS video stabilisation tools and provide guidelines for their use in videos with different flight and ground conditions.

An innovative approach is presented to optimise image-velocimetry performances for surface flow velocity estimates (and thus remotely sensed river discharges). Synthetic images were generated under different tracer characteristics using a numerical approach. Based on the results, the Seeding Distribution Index was introduced as a descriptor of the optimal portion of the video to analyse. A field case study was considered as a proof of concept of the proposed framework showing error reductions.

A new comprehensive dataset of annual maximum rainfall depths recorded in 1 to 24 consecutive hours in Italy is presented. More than 4500 stations are considered, spanning the period between 1916 and 2014. For the first time, a national dataset of annual maxima allows for an updated characterization of rainstorms in Italy, getting rid of the regional borders. The spatio-temporal analyses of the highest rainstorms represent a robust starting point for the assessment of the extreme rainfall risk.