According to the National Disaster Management Authority (NDMA), one of Pakistan's worst floods, triggered by exceptionally heavy monsoon rains, affected around 33 million people nationwide. During floods, it’s necessary to gather observations from the disaster site, and these observations are used to create flood-level maps to aid in emergency operations. People in flooded areas shared text and images on Social media platforms to show the current situation. In this paper, we suggested a mask R-CNN model-based approach to predict flood water levels on images collected from social media. We face these challenges i) the size of the object that appears in the image is unknown ii) There may be variations in the height of the flood water that appears in various parts of the image, and iii) the objects that may be submerged in water are partially visible. We addressed these issues by identifying class objects with known sizes to estimate water levels. We train the Mask R-CNN model on the flood-water dataset and then finally check the ability of the train model on collected images from real-time flood areas. The Mask-RCNN model achieved 0.85 accuracy in detecting submerged objects with an error margin of just 0.15 cm in water level estimation. Keyword: Deep Learning, CNN, Mask R-CCN, Flood level Estimation, Social media Image Recognition