Centroid-based endmember optimization of the triangular space method for fractional cover estimation: Mapping fractional cover of a vegetated ecosystem on Sentinel-3 OLCI image

Accurately estimating fractional cover of vegetated ecosystems over large areas is essential for many scientific studies, including climate change, land cover and land use, etc. Taking both accuracy and large spatial coverage into account, different methods have been proposed, such as upscaling from high to low spatial resolution remote sensing images, and harmonized data from varied sources. In this work, a new method, called centroid-based endmember optimization (CEO), is proposed to assist endmember selection for fractional cover estimation using the triangular space method. The basic idea is to retrieve endmembers from a triangular space built on a fine spatial resolution image, then correct and apply them to a coarse spatial resolution image. The method is discussed and tested using Sentinel-2 MSI and Sentinel-3 OLCI images to estimate non-photosynthetic vegetation (NPV), photosynthetic vegetation (PV), and bare soil (BS) fractional cover. With CEO, the fractional cover of an OLCI image can be estimated more accurately, reducing the work of mosaicking MSI images to acquire fractional cover of the same large area. The premises that CEO can be applied effectively are: (1) the acquisition dates of the MSI and OLCI images are close, ensuring a similar land cover; and (2) the spatial overlap between the MSI and OLCI images covers enough NPV, PV, and BS endmembers. When taken the average fractional cover retrieved from an MSI image as truth value, the CEO method reduced the estimation difference to 0.7%, compared to the differences of 8.1% and 6.6% retrieved using uncorrected or incompletely corrected triangular space method of an OLCI image, respectively. In addition to the average fractional cover estimation, the histogram of fractional cover distribution also improved obviously. When applying CEO to a full OLCI image, two full MSI images covering different locations were used for fractional cover validation, which supported a robust estimation result.