Mapping dynamics of large-scale high-precision pond datasets using a semi-automated method based on deep learning

Abstract

With their large numbers and widespread distribution, ponds are crucial in stormwater interception, biodiversity, and freshwater resource conservation. However, due to their small size and shallow depth, ponds are highly susceptible to anthropogenic activities and climate variability, making it necessary to map their numbers, distribution, and change dynamics. Relying only on deep learning (DL) techniques is insufficient to create a pond identification dataset that does not contain errors. This study is the first of its kind that proposes a workflow to identify small ponds (<5 ha) with minimal errors from the National Agricultural Imagery Program (NAIP) high-resolution aerial imagery using the combination of DL and a manual cross-correction approach. Ponds in South Carolina, United States, were detected and delineated for 2017 and 2019 using U-Net models. Next, the detection results from both years were used as reference data for cross-correction, removing false detections and adding omissions to obtain the refined high-precision pond datasets. The pond datasets were compared to the existing public datasets (JRC and NWI) to evaluate the performance of the proposed method. Finally, changes in ponds between two years and the predominant land cover around each pond in 2019 were analyzed in our study. The results showed that the refined high-precision pond dataset containing 70,449 ponds in 2017 and 71,858 ponds in 2019, with an average size of 0.5 ha, fills an important gap in existing pond data. The existing public datasets (JRC and NWI) do not identify 61.72% and 41.03% of the new high-precision pond dataset developed as part of our study in 2019. Based on land cover data, the largest number of ponds were located in forested areas (23,188 ponds, 0.76 ponds/km²), followed by wetlands (15,782 ponds, 0.76 ponds/km²). In contrast, barren land and hay/pasture had the highest pond density, reaching 2.67 ponds/km² and 1.93 ponds/km². A total of 2,979 ponds experienced changes between 2017 and 2019, and 69,664 ponds remained unchanged. The types of pond changes can be categorized as new pond construction, water level changes, and pond disappearance. Our study significantly advances a workflow and method for pond detection that leverages deep learning over large areas in diverse ecological regions and can provide high-precision pond datasets with minimal errors for pond evaluation and management.