Development of river water level estimation from surveillance cameras for flood monitoring system using deep learning techniques

Around 70% of global disasters are related to hydro-meteorological events such as drought, floods, and cyclones. Therefore, researchers and experts carried out many studies on flood hazards in order to reduce the impact of flood magnitude and flood frequency. In Malaysia, a telemetric forecasting system is currently been used in flood monitoring systems. However, data information obtained from this system is one spatial dimension and one point-based station, thus it cannot represent the dynamics of the surface water extent. Therefore, this study introduces a visual surveillance concept to monitor the flood event in a specific area, based on surveillance cameras and computer vision approaches to obtain instant flood inundation information during flood events. A deep learning approach was proposed for water segmentation so that it can be applied to various water scenarios and backgrounds. However, conventional image segmentation techniques were also carried out to ensure the usage of deep learning is worth it. The conventional segmentation methods used in this work are thresholding, region growing, and hybrid technique known as GeoRegion. The findings demonstrated that these methods are handcrafted and the algorithms need to be changed when applying to different images, which is not practical to be used during flood disasters. Hence, deep learning technique was chosen for water segmentation procedure in this work. Two different networks were applied in this study, namely DeepLabv3+ and SegNet, for detecting water regions before estimating water levels from surveillance images. Water level estimation was predicted based on the elevations from LiDAR data. Based on the experimental results, it was found that the DeepLabv3+ network performed better than the SegNet network by achieving above 93% for overall accuracy and IoU metrics, and approximately 82% for boundary F1 score (BF score). The Spearman’s rank correlation obtained between water level measured by the sensor and water level estimated from the proposed framework was 0.92 which indicates a strong relationship. By integrating the estimated water level with a 3D model developed from LiDAR data, flood simulation was performed. Besides, volume of water was also computed from the 3D model. The findings demonstrate that the water volume increased as water level increased. Lastly, a graphical user interface was developed for water segmentation and water level estimation analysis that could be applied during the flood events. Hence, the proposed work can help in improving the current monitoring and emergency warning abilities against flood events, serving as a complement to the currently used quantitative precipitation forecasts and in-situ water-level measurements.

Text NUR ‘ATIRAH BINTI MUHADI -IR.pdf Download (1MB)

Actions (login required)

View Item