Updating forest maps through data assimilation using remotely sensed data and regression methods designed to avoid bias trends

We applied data assimilation for updating map information about growing stock volume in forests in a study area in northern Sweden, across a study period ranging from 2010 to 2022. Novel features of the study were that (i) we applied newly developed regression techniques for predicting forest characteristics from remotely sensed data, designed to avoid the bias trends that arise from standard regression methods, (ii) we applied growth models that utilized information about site quality and age, assessed wall-to-wall for the study area, based on data from repeated airborne laser scanning surveys, and (iii) we used a fully empirical approach to computing weights for the DA filter. The results showed that the accuracy obtained for predictions of growing stock volume from an initial laser scanning survey could be improved upon across the study period when a sequence of predictions using optical satellite data, digital aerial photos, and a final laser scanning survey were assimilated. Using only a sequence of optical satellite data and digital aerial photos, the accuracy of the initial laser scanning-based predictions could be maintained across the study period. The new regression methods performed better than standard regression methods in terms of avoiding bias trends, but the best overall results in terms of accuracy were obtained for standard regression combined with classical calibration. The study confirms findings from previous similar studies that data assimilation has a potential to maintain or slightly improve the accuracy of growing stock volume predictions from an initial high-quality laser scanning survey through assimilating a series of predictions from lower-quality remotely sensed data across a relatively long period of time.