Lang, M., Gulbe, L., Traškovs, A. and Stepčenko, A. 2016. Assessment of Different Estimation Algorithms and Remote Sensing Data Sources for Regional Level Wood Volume Mapping in Hemiboreal Mixed Forests. Baltic Forestry 22(2): 283-296.

   Remote sensing data provide opportunity to estimate wood volume in vast areas with lower financial expenses compared to field measurements. In this study, we tested wood volume mapping of hemiboreal mixed forests at stand-level and regionally using forest management inventory data as a reference set, various remote sensing data sources (Landsat-5 TM, Landsat-7 ETM+, SPOT-4 HRVIR, ALOS PALSAR, airborne laser scanner data) and three nonparametric estimation algorithms (k-nearest neighbours (kNN), general regression neural network, regression tree). The experiment in Kurzeme region, Latvia, was organized as case studies regarding some aspects of the estimation procedure: impact of randomness in reference set sample on the k-NN volume estimation, assessment of the influence of the image and training plot combination on the k-NN volume estimations, comparison of the estimation algorithms and comparison of multisource and multitemporal data fusion. All the estimators performed quite similarly due to the complex relationships between forest inventory data and remote sensing data. The smallest RMSE=60 m3/ha was achieved in the special study site in Slitere National Park by combining five feature variables that included the 70th percentile of the ALS point cloud height distribution, green band from the Landsat and SPOT image, and NIR and SWIR bands from the Landsat image. When spectral feature variables and reference samples from full-size satellite scenes were used, the RMSE of wood volume estimates ranged from 72 m3/ha to 129 m3/ha for forest in the scenes. Higher estimation accuracy was obtained with mid-growing season Landsat images and then with SPOT images from the snow-covered period. Case studies indicated that the estimation accuracy depends on a particular image, but the randomness in the reference set does not impact accuracy substantially when there is a sufficient number of reference sample plots. The combined influence of a particular image and reference samples for the image was detectable and the RMSE of the stand-level wood volume estimates in the image overlap areas ranged between 17 m3/ha and 42 m3/ha, and mean error of estimate ranged from - 26 m3/ha to 21 m3/ha.

Keywords: wood volume, multisource remote sensing data, nonparametric estimators, forest management inventory.