Automatized Sentinel-2 mosaicking for large area forest mapping

Abstract

Creating maps of forest inventory variables is commonly taking advantage of satellite images, which are mosaicked together for gaining larger coverage. Recently, mosaicking has increasingly shifted towards user friendly cloud-based online environments such as Google Earth Engine (GEE), which are equipped with huge image repositories and extensive processing capabilities. This enables the easy transferability of workflows into new image sets and diversifies the range of methodological options for mosaicking. The quality control of the output mosaic, ensuring that the reflectance values are representative to the targeted land cover, is however primarily based on certain assumptions or pre-set rules which may not always produce an optimal result. Our study focuses on assessing and comparing the performance of three different mosaicking algorithms for predicting forest inventory variables, based on an extensive set of field data on the main site type, fertility class, and volume and biomass of growing stock. One of the compared mosaics derives from manual image selection, thus enabling rigorous visual quality control, and two others are resting on GEE-assisted automatized methods which include applying a percentile-based statistic over all the input reflectance values and selecting the best pixels using predefined quality indicators. The results indicate that the manual and the percentile-based mosaics are generally providing the best and relatively equal performance levels. Compared to them, the quality-based mosaic has slightly lower accuracy particularly when predicting continuous variables (i.e., the volume and biomass of growing stock) and it suffers from minor image defects. For the total volume of growing stock, for example, the RMS errors are 56.22 % for the manual, 56.33 % for the percentile-based, and 59.47 % for the quality-based mosaics, respectively. These results indicate that from the perspective of large area forest mapping, automatically generated mosaics may provide approximately similar accuracy as compared to manually controlled workflow at a fraction of the workload.