I. Introduction
Outfalls are the last line of defense for pollution sources entering the ocean, rivers, and lakes, and are related to the safety of the water environment and the health of aquatic organisms [1]. It is essential to determine the distribution location and number of outfalls for designated waters in a timely and accurate manner. Recently, studies on outfalls have attracted a lot of attention from environmental protection authorities in various countries [2], [3]. China conducted comprehensive outfall surveys of the country's rivers and lakes in 2011, and identified the distribution location and number of outfalls [1]. Starting in 2019, the Ministry of Ecology and Environment of the People's Republic of China began an outfall survey on the Yangtze River, Yellow River and Bohai Sea. They planned to complete the survey on all of the outfalls in seven river basins and offshore which include the Yangtze River, Yellow River, Huai River, Hai River, Pearl River, Songliao and Taihu Lake by the end of 2025. In order to standardize and unify the means of the aforementioned study and improve its efficiency, the Ministry of Ecology and Environment of the People's Republic of China adopted a combination of satellite remote sensing technology, Unmanned aerial vehicle (UAV) remote sensing technology, and manual field survey. They divided the outfall survey into three levels to carry out the work. The first level detects suspected outfall targets by means of remote sensing interpretation. The second level confirms the suspected targets obtained from the first level by means of manual field survey and supplements the outfalls that are missed. The third level reviews the problematic outfalls and carries out precise remediation of some important areas. Researchers have conducted some valuable studies on the first level of survey methodology. Satellite remote sensing technology acquires ground images from an overhead perspective and can be used for outfall survey work. Zhang et al. [4] combined web crawler technology with satellite remote sensing interpretation technology to determine the location of industrial outfalls. This method obtained a large amount of industrial information and location data from the Internet, and then combined the remote sensing interpretation technique to detect the specific location of the outfalls. But the method heavily relies on the data quality of satellite images. Trinh et al. [5] used remote sensing inversion to study the effects of wastewater on the offshore environment. It used remote sensing data such as Landsat 8 TIRS and OLI to invert sea surface temperature and Chlorophyll-a concentration in the offshore area, and then analyzed to obtain the impact of wastewater on the offshore environment. Finally, the outfall conditions were determined. Gierach et al. [6] used multi-source remote sensing data to invert the sea surface roughness, sea surface temperature, and Chlorophyll-a concentration in the offshore. There are often differences between polluted and normal water bodies. It obtained the outfall conditions based on this principle. Similarly, Nezlin et al. [7] used multi-source remote sensing data to invert chlorophyll concentrations, etc., and thus indirectly identify the outfalls. Wu et al. [8] proposed a multi-scale remote sensing detection method. It used medium-resolution satellite images for water quality parameter inversion, and then used high-resolution satellite images for outfall detection. These methods rely heavily on the accuracy of parameter inversion and cannot effectively detect outfalls with hidden discharges. Using satellite remote sensing technology to investigate outfalls has many challenges. Compared with satellite remote sensing technology, airborne remote sensing technology can provide both spatial resolution and temporal resolution with decent qualities. However, it is costly to operate and can be affected by factors such as airspace policy and weather conditions. The UAV remote sensing technology works in a more flexible and lower cost way, and can provide high spatial resolution ground images in a near-ground-up manner, which is suitable for outfall investigation work. Relatively little work has been done in this area. Huang et al. [1] proposed a geographic knowledge-based outfall detection method that combines digital surface models (DSM), river data, and UAV remote sensing images for outfall detection. This method allows for accurate detection of outfalls. But a large amount of auxiliary data, such as DSM of watersheds, river data, etc., is required. Qi et al. [9] has developed an outfall identification system that can automatically detect images taken by UAVs. This solution can only detect one image at a time, which is inefficient. Please note that both of the above approaches are unable to perform real-time detection and cannot provide timely feedback of survey results to staff. And the datasets for outfalls are quit rare. Most of previous studies were based on the internet images.