Remote sensing image extraction for coastal aquaculture ponds in the Guangxi Beibu Gulf based on Google Earth Engine

Abstract: Aquaculture ponds have expanded significantly with the increase of food demand under an ever-increasing population, especially in the coastal areas of China. However, a large number of aquaculture ponds have posed a great threat to the ecological environment, such as the destruction of coastal wetland, together with the water and soil pollution, although economic benefits have been gained during this time. Therefore, the spatial distribution of aquaculture ponds is of great significance for the scientific management of coastal zones and the sustainable development of fishery. There are many bays, lagoons, and coastal marshes in the Beibu Gulf of the Guangxi Zhuang Autonomous Region in China, particularly where the coastal terrain is more complex. Moreover, there is some competition between aquaculture land and other land-use types, leading to the fragmentation of aquaculture land patches. Furthermore, there is also limit utilization of traditional remote sensing to identify the aquaculture ponds in this area. In this study, taking the Beibu Gulf of Guangxi province in China as the study area, a new remote sensing identification was proposed using Google Earth Engine (GEE) platform and time-series remote sensing data. Firstly, all Sentinel-1 and Sentinel-2 remote sensing datasets were collected in 2019. The following classification eigenvalues were constructed that: 1) Inundation frequency (IF) was used to evaluate the water bodies at pixel scale. 2) The annual mean value of pixels SWIR2 and VH were calculated to eliminate the water identification error caused by building shadow and impervious water surface. 3) Image collection of NDWI time-series was integrated to calculate the mean value between 85%～95% in an ascending order at the pixel level, particularly for better identification of dikes between ponds. The optimal segmentation threshold was then determined to extract the aquaculture ponds using a large number of training books. Finally, the object-oriented method was used to screen out the objects for better classification effects. The segmentation was also carried out again to improve the recognition. The results showed that the total area of aquaculture ponds was 199.3km2, covering the entire coastal area. Furthermore, the overall accuracy of identification was 0.921, and the Kappa coefficient was 0.842. Four regions were selected as examples to visually interpret the aquaculture ponds using the Google high-resolution images, thereby further evaluating the validity of identification. The identified area accounted for more than 90% of the area of visual interpretation in the large-scale concentrated aquaculture ponds, while 80.76% of the area of densely arranged small ponds, compared with the visual interpretation. The obtained values were closer to the actual aquaculture surface area. It infers that the temporal remote sensing data can widely be expected to serve as the classification characteristic value, while effectively exclude the abandoned ponds, paddy fields, and seasonal waters. A better performance was achieved in a large-scale complex environment, particularly on higher efficiency of remote sensing identification and extraction of aquaculture ponds. This finding can provide sound technical support to the AI monitoring the ecological environment and spatial optimization of coastal areas in modern aquaculture.