Contribution of ECOSTRESS thermal imagery to wetland mapping: Application to heathland ecosystems

Abstract

While wetlands have been extensively studied using optical and radar satellite imagery, thermal imagery has been used less often due its low spatial – temporal resolutions and challenges for emissivity estimation. Since 2018, spaceborne thermal imagery has gained interest due to the availability of ECOSTRESS data, which are acquired at 70 m spatial resolution and a 3–5 revisit time. This study aimed at comparing the contribution of ECOSTRESS time-series to wetland mapping to that of other thermal time-series (i.e., Landsat-TIRS, ASTER-TIR), Sentinel-1 SAR and Sentinel-2 optical satellite time-series, and topographical variables derived from satellite data. The study was applied to a 209 km² heathland site in north-western France that includes riverine, slope, and flat wetlands. The method used consisted of four steps: (i) four-year time-series (2019–2022) were aggregated into dense annual time-series; (ii) the temporal dimension was reduced using functional principal component analysis (FPCA); (iii) the most discriminating components of the FPCA were selected based on recursive feature elimination; and (iv) the contribution of each sensor time-series to wetland mapping was assessed based on the accuracy of a random forest model trained and tested using reference field data. The results indicated that an ECOSTRESS time-series that combined day and night acquisitions was more accurate (overall F1-score: 0.71) than Landsat-TIRS and ASTER-TIR time-series (overall F1-score: 0.40–0.62). A combination of ECOSTRESS thermal images, Sentinel-2 optical images, Sentinel-1 SAR images, and topographical variables outperformed the sensor-specific accuracies (overall F1-score: 0.87), highlighting the synergy of thermal, optical, and topographical data for wetland mapping.