Combining UAV and Sentinel-2 auxiliary data for forest growing stock volume estimation through hierarchical model-based inference

Remotely sensed (RS) data are becoming increasingly important as sources of auxiliary information in forest resource assessments. Data from several satellites providing moderate image resolution are freely available (e.g. Sentinel-2). In addition, very-high-resolution three-dimensional data are available due to the advent of unmanned aerial vehicles (UAV). The increasing availability of auxiliary data offers new opportunities for largescale forest surveys using UAVs. A recently developed hierarchical model-based mode of inference makes it possible to use hierarchically nested auxiliary data in estimating population properties, such as total or mean biomass or volume, and their corresponding uncertainties in a statistically appropriate manner. In this study, hierarchical model-based inference was used to estimate growing stock volume (GSV; m(3) ha(-1)) and its variance using a small sample of field data, a larger sample of UAV data, and wall-to-wall Sentinel-2 data in a study area in SE Norway. The main objective of the study was to compare the performance, in terms of precision, of hierarchical model-based inference (denoted Case C) against two alternative cases. These were (1) model-based inference based on field data and wall-to-wall data, collected either with airborne laser scanning (Case A.1) or Sentinel-2 data (Case A.2), and (2) hybrid inference using a small sample of field data and a larger sample of UAV data (Case B). A second objective was to assess the possibility of reducing the UAV sampling intensity when adopting Case C rather than 13, without decreasing the precision of the GSV estimates. The results, calculated as standard error as percentage of the mean ((SE) over cap (%)), indicated that in case C the precision was of similar magnitude ((SE) over cap (%) = 3.44%) as for Case A.1 ((SE) over cap (%) = 3.69%) and for Case B ((SE) over cap (%) = 3.58%). The standard error of Case A.2 was nearly twice as large ((SE) over cap (%) = 5.81%) as the rest of the cases. The results also indicated possibilities of reducing the UAV sampling intensity without losing precision in cases where wall-to-wall Sentinel-2 data are available (Case C). The same precision for Case C with only five UAV samples was achieved as for Case B with 55 UAV samples. Thus, the study highlights the cost-efficiency of applications of UAV as in Case C and also provides first insights in the use of Sentinel-2 data for GSV estimation in boreal conditions.