Visual Object Recognition in Water Resources - Developing a Semantic Segmented Water-related Object Dataset

The application of deep learning in water resources engineering has been grown over the past few years. One of the main capabilities of the ConvNets is the semantic segmentation and interpretation of visual scenes to identify distinct objects in images and videos. In water resource management, segmentation of waterbodies along with other water-related objects offers a new heterogeneous source of information with extensive potential for various applications, such as flow rate estimation, drought management, and real-time flood mapping. The first step for such applications is developing an extensive water-related object dataset by which ConvNets could be trained. This study presents the formation of the first exclusive semantic segmented water-related dataset, which consist of 5,000 pixel-wise annotated images with 51 different labels. The dataset contains 18 types of natural water system labels, such as river, sea and wetland, and 17 built water-related objects, including dam, culvert, canal, etc. In addition, 21 other objects, such as person, car, road, vegetation, that supports contextual reasoning for image recognition are annotated. Finally, sets of existing benchmark ConvNets in the field of computer vision are trained by this dataset and the performance of each model is reported using the confusion matrix.