Assessing the transferability of airborne laser scanning based models for predicting natural forest attributes

Abstract

Natural forest areas which have usually been unmanaged already several decades are often small and isolated from one another. The performance of airborne laser scanning (ALS) based management forest inventory in natural forests is unknown, since ALS inventories are typically based on field reference data from managed forests. The aims of the study were to (1) Develop ALS based linear regression models for predicting forest attributes in natural forests, and (2) Assess how reliably such models can be applied across different inventory areas. The forest attributes considered were growing stock volume (m³ ha^–1), volume of standing dead wood (m³ ha^–1), number of stems (ha^–1), dominant height (m), and structural complexity by Gini coefficient (0–1).

The inventory areas were located in Southern and Eastern Finland. Data from four inventory areas were used in model development, and sample plots measured from even-aged forest area and continuous-cover forest area were used for validation. The point densities of ALS data in the inventory areas varied between 4 and 13 m^–2, and the number of plots in the field datasets ranged from 14 to 148. Linear regression models were fitted with field measured plot attributes as responses and two ALS features as predictors. A total of 25 forest variable models were constructed, and 150 model predictions were made. Analyses were conducted separately for each natural forest data set, but general models fusing all natural forest data sets were also constructed. The models' performances in transfers across inventory areas were quantified with relative root mean square error (rRMSE) and bias rBias values.

The accuracies of the models in their fitting datasets were in line with the previous studies. In terms of the transferability, the performance of the models was modest. On average, the rRMSE doubled and rBias increased considerably in the transfer. On average, the rRMSE increased by 9% for growing stock volume, 190% for dead wood volume, 31% for number of stems, 6% for dominant height, and 9% for Gini coefficient. If the obtained accuracy of growing stock attributes were acceptable for natural forest inventory, one could rely on model transfer or the use of general models.

Keywords: primary forest; remote sensing; airborne laser scanning; forest variable; area-based approach; regression model