Epsilon Archive for Student Projects

Abstract

Reliable forest estimates are a very important aspect of sustainable forest ecosystems management, but commonly difficult to obtain for large areas. However, with ALS-based forest inventory it is possible to obtain significant estimates. This study tested the application of models currently used to estimate forest variables such as basal area weighted mean height (H), basal area weighted mean diameter (D), basal area (BA) and volume (V) at plot level for the whole region of Götaland, a part of Sweden. The models were adjusted by the combination of both laser and NFI field data in order to compare the effect of the use of: 1) two different training datasets (local and regional) and; 2) two types of NFI plots (temporary and permanent). The analyses were done by linear regression. The accuracy evaluation was conducted by means of absolute and relative root mean square errors (RMSE) and bias through cross-validation on plot level and application of the estimations on stand level, which consisted of 65 stands within the study area. Each with an average size of approximately 8 ha. The assessment results on stand level of the two different datasets for all variables validated (H, D, BA and V) show RMSE and bias slightly better for the local models (8.8%, 11.8%, 14.0%, and 15.9% respectively), with the highest increase in RMSE for the estimated volume (3%) when applying the regional model (9.7%, 13.8%; 14.7% and 18.8% respectively). Moreover, the models calibrated with only permanent plots performed better than the models calibrated with both permanent and temporary plots. In a second stage of the evaluation, explanatory dummy variables related to geographic position of the forest such as if the forest were located close to the coast, and a variable created by the combination of altitude and latitude were tested. With the dummy variables, NFI plots from the buffer zones of 20 and 50 km around the coast, a slightly lower RMSE for D and BA were obtained. That indicates that the proximity to the coast has impact on the site characteristics of the stand; therefore, the regression models may operate differently according to the forest type. The analysis implies that the models calibrated with the regional training dataset produced results comparable to the ones produced using models calibrated with local training datasets.