当前位置：
X-MOL 学术
›
Remote Sens. Environ.
›
论文详情
Our official English website, www.x-mol.net, welcomes your
feedback! (Note: you will need to create a separate account there.)
Inventorying ponds through novel size-adaptive object mapping using Sentinel-1/2 time series
Remote Sensing of Environment ( IF 11.1 ) Pub Date : 2024-10-30 , DOI: 10.1016/j.rse.2024.114484 Denghong Liu, Xiaolin Zhu, Meredith Holgerson, Sheel Bansal, Xiangtao Xu
Remote Sensing of Environment ( IF 11.1 ) Pub Date : 2024-10-30 , DOI: 10.1016/j.rse.2024.114484 Denghong Liu, Xiaolin Zhu, Meredith Holgerson, Sheel Bansal, Xiangtao Xu
Ponds are an important source of greenhouse gases (GHGs) to the atmosphere, yet evaluating their role in global biogeochemical cycling is currently hampered by limitations in quantifying their global distribution. Existing satellite-derived estimates of lake distributions have difficulty identifying small lakes (5–10 ha) and ponds (<5 ha) due to limitations in satellite resolution and challenges extracting individual small waterbodies from low-albedo surfaces, vegetated water, and lotic water systems including rivers and streams. In this study, we developed generalizable pond mapping strategies based on their spatial-temporal-spectral characteristics to fully exploit accessible medium-resolution optical and synthetic aperture radar (SAR) time series to identify ponds. Our novel approach entails: (1) making full use of ponds' characteristics from an object-based perspective; (2) extracting pond objects using seeds of prominent water pixels defined by the SAR VH signal; (3) constructing training samples of ponds with high representativeness; and (4) improving inter-class discrimination by combining features from optical and SAR data. We designed a novel Optical-SAR Pond Object Mapper (OptiSAR-POM) to achieve an improved estimate of pond size distribution by incorporating mapping strategies into the object-based image analysis framework. We generated landscape objects through an elaborate water-focused segmentation approach, which adaptively aligned the segmentation parameters with the size and distribution patterns of ponds to identify small waterbodies and increase inter-class variability. We further introduced an interactive learning process to construct random forests for object-based classification, which incorporated adaptive empirical thresholds to identify potential pond objects and select representative training samples of varying sizes. We tested the OptiSAR-POM framework using Sentinel-1/2 time series at three county-level study sites and three supplementary watershed-level study sites in the United States and China. Our approach yielded high overall accuracy (>95 %) for all sites and highlighted the ability of Sentinel-1/2 imagery to accurately detect small ponds (0.1–1 ha) across diverse landscapes. The average producer's accuracy for small ponds at county-level sites improved by ∼45 % compared to that of all other products with a 10-m or higher spatial resolution, addressing the absence of such information in existing regional and global datasets. The generated county-level pond maps revealed the numerical dominance of ponds in lentic waters, their substantial area contribution in human-impacted regions, and the relevance of studying biogeochemical processes in smaller waterbodies.
更新日期：2024-10-30
京公网安备 11010802027423号