Improving the accuracy of timber volume and basal area prediction in heterogeneously structured and mixed forests by automated co-registration of forest inventory plots and remote sensing data

Accurate georeferencing is essential if forest indicators, such as timber volume, are to be modelled and predicted area-wide by establishing a linkage between local inventories and remote sensing data. Nevertheless, due to inaccuracies in global navigation satellite system (GNSS) measurements under a closed canopy, determining the exact position of sample-plots in a forest inventory is a major challenge. In this study different methods were evaluated, each of which is designed to improve the co-registration between field measurements and remote sensing data in a forest inventory. The methods were evaluated in two areas in Switzerland (Bremgarten and Zurich), all of which have heterogeneously structured and mixed forests. A simple algorithm that searches for the best match between tree-tops, detected in remote sensing data, and an inventory tree-top point-cloud led to unsatisfactory results. The failure of the algorithm was primarily related to the lack of accurate single-tree-identification (STI) methods in airborne laser scanning (ALS) data (10 points m(-2)) on deciduous and mixed forest stands. These inaccuracies hampered a successful co-registration of the two data sources. To omit the single-tree-identification (STI) step, methods that rely on the comparison between an artificial canopy height model (CHM) calculated from the inventory data and the CHM generated from ALS data were tested. Correlating the two CHMs made it possible to identify plausible positions within the search range. The quality of all co-registration methods was assessed by the leave-one-out cross-validated root-mean-squared error (RMSE) of the timber volume estimate of the subsequently calibrated regression model. The best results were achieved with a method that modelled inventory tree-crowns as spheres and that applied a correlation metric named SQDIFF_-NORMED. Co-registration made it possible to increase the model accuracy of timber volume estimates in Bremgarten by 31 RMSE-%, i.e. from 134.4 m(3) ha(-1) (without co-registration) to 92.6 m(3) ha(-1) with co-registered positions. Additional testing of the identified superior method with a larger inventory dataset of the canton of Zurich, Switzerland confirmed these results. There, the RMSE of the basal-area estimate was improved by 10 RMSE-%, from 13.56 m(2) ha(-1) to 12.18 m(2) ha(-1). For these CHM-based methods, integrating the information from a deciduous-evergreen (DecEv) raster improved the positional accuracy but not the overall predictive power of the regression models.

Automated summary

Accurate georeferencing is crucial for modeling and predicting forest indicators like timber volume. Different methods were evaluated in this study to improve the co-registration between field measurements and remote sensing data in a forest inventory. The best results were achieved with a method that models inventory tree-crowns as spheres and uses a correlation metric called SQDIFF_-NORMED.