Enhanced Urban Flood Monitoring: Integrating Advanced Semantic Segmentation and Human Facial Feature and Posture Analysis
Bhavana B. Nair, P. Vallimeena, Uma Gopalakrishnan, Sethuraman N Rao, Shivsubramani Krishnamoorthy · IEEE Access · 2024
In recent years, flooding has become one of the most frequent life-threatening natural disasters. As smartphone cameras are user-friendly and easily accessible, many images are captured and posted on social media during events like a flood. The objects in those images can be used as a reference to estimate floodwater depth, suggesting appropriate processes and equipment needed for rescue operations. Most often, those images contain humans fully or partially submerged in floodwater in different poses. A standing human pose is considered to be ideal for accurate water depth estimation. One of the challenges when using humans as reference is accurately predicting their pose to avoid false positives. Firstly, a valid image identification subsystem, consisting of a CNN-based algorithm for human facial features and posture analysis, was designed to tag a flood scene image as invalid and discard it. Typical cases are ‘a child over the shoulder’, ‘a child on the waist’, ‘a human on a vehicle (two-wheeler or boat)’, ‘a human sitting over elevation’ and ‘a human leaning over or swimming’; achieving accuracy rates of 71.43%, 72.86%, 81.43%, 68.57% and 61.43% respectively. Secondly, a new crowd-sourced image dataset, named AMRITAFLOODDATA, of 10,854 flood scene images from the web and social media was created. Thirdly, these images were annotated and used to train a CNN-based semantic segmentation model with a validation accuracy of 93.2%. Lastly, as an additional data-point, Z-factor is calculated and also, the average height of the reference human in the flood scene is identified based on gender, age-group and ethnicity classification. The accurate segmentation of water and human in an image led to the precise detection of the waterline, enabling the estimation of flood water depth with an error range of ±20 cm.