Linear Risk Source Of Drinking Water Source Extraction Research Based On Object-based Deep Learning

Water pollution is an increasing problem with the rapid development of society.Water quality security of drinking water sources is closely related to urban operation and people’s life and health.Roads and bridges are the carriers of land transportation mobile source(such as tankers and vehicles operating chemicals)and prone to water pollution.Accurately extracting roads and bridges by remote sensing technique is of great significance for environmental monitoring of drinking water source.Roads and bridges are strip-shaped and are interrelated in space.In this paper,they are collectively referred as "linear risk source" and extracted by the techniques of deep learning and object-based image analysis.The main works are as follows:(1)A road network extraction method is proposed based on deep learning semantic segmentation.A sample collection approach is first designed to train model.Then mainstream semantic segmentation models,such as U-Net,PSP Net and Deep Lab V3+,are compared and analyzed.The U-Net model with the best road extraction effect is finally selected for preliminary road network extraction.(2)A heuristic searching method integrating depth probability is proposed for accurate road network extraction.Calculate the skeleton lines of the road region primitives,and remove non-target information such as burrs and building adhesion to obtain road linear primitives.The heuristic function is then constructed based on the deep learning prediction probability and the skeleton direction to extract road accurately.(3)A bridge extraction method using shape and spatial relationship is proposed.Water body is extracted using deep learning method.Bridge extraction rules considering shape characteristics and spatial relationship with the water body are established for preliminary bridge extraction.Finally,refine the bridge extraction results by the bridge skeleton calculation and improve extraction accuracy.In this paper,the drinking water sources in different regions of Nanjing city are taken as experimental areas.The experimental results show that the proposed methods can effectively extract linear risk sources and improve the extraction accuracy.