A Self-Tuning Algorithm for Compositing Clear-Sky Reflectance Data at 250-Meter Spatial Resolution for the Fengyun-3D MERSI-II Land Bands Over China

The MEdium Resolution Spectral Imager-II (MERSI-II) on board the Fengyun-3D satellite is an advanced imaging instrument with 250/1000 m spatial resolution. This instrument is a valuable resource for terrestrial remote sensing, such as clear-sky surface reflectance products. Clear sky composites are typically generated using fixed and empirical threshold values based on specific land surface type, climate, and weather conditions. Moreover, existing methods usually require cloud masks, which can be inaccurate and affect clear sky composites. We took inspiration from natural language processing and viewed the time-series reflectance data as per-pixel temporal sequences. A new self-tuning approach transforms the temporal reflectance data to contextualized embeddings in a high-dimensional feature space through pretraining, followed by a classification layer that is fine-tuned by a few labeled data to identify scene ID mask (cloud shadow, clear sky, thin cloud, and thick cloud). We then proposed a compositing strategy that takes into account the spatial connectivity and temporal consistency of clear-sky pixels. Comprehensive evaluations, including visual assessment, annual temporal profiles of NDVI/NDWI at a site scale, and NDVI spatial distribution at a national scale, of the proposed approach applied on 250 m bands were conducted by comparing it with four typical compositing methods (CCRS-BH, CCRS-FH, MaxNDVI, and MinRed), as well as the MODIS clear sky products. The proposed approach outperforms the four typical compositing methods and achieves a similar product level to MODIS composited products but with a higher spatial resolution.