Silvilaser 2019 - Poster Presentations »
Predicting species-specific diameter distributions in boreal forests using bi-temporal ALS, multispectral ALS, and aerial images
A diameter distribution describes the size distributions of trees in a forest and is a vital output from a forest inventory. Diameter distributions are predicted using remotely sensed data in modern forest inventories. Unispectral airborne laser scanning (ALS) data are the most commonly used remotely sensed data from which to derive features for the prediction of diameter distributions. Typically, the spectral features derived from aerial images are also used when predictions are carried out by tree species. Nowadays, remotely sensed data are increasingly available, and datasets, such as multispectral and bi-temporal ALS data, are also accessible. It is evident that studies are needed to efficiently utilize different data sources in the modern ALS-based forest inventories.
Species-specific diameter distributions are needed in Finland. Because of the low number of commercial tree species, the prediction by tree species is possible. A nearest neighbor (NN) approach enables to predict forest attributes and diameter distributions simultaneously. We applied the NN approach to predict diameter distributions by tree species in a study area that is located in eastern Finland. The study area represents typical managed boreal forests where Scots pine and Norway spruce dominate. We used a separate training (n = 424) and validation dataset (n = 420). The predictor variables for NN approach were selected using an optimization-based feature selection. Our objective was to examine how the different combinations of remotely sensed data affect to the predictive performance associated with the predicted diameter distributions. We used the following remotely sensed datasets: unispectral ALS data (leaf-off and leaf-on), multispectral ALS data (leaf-on), old unispectral ALS data (leaf-off), and aerial images. The predictive performance was evaluated by means of RMSE% and BIAS% values associated with the predicted timber assortment volumes (logwood and pulpwood).
The combination of ALS data and aerial images was selected as a reference combination since it is operationally used in Finland. The results showed that the reference combination of features outperforms the multispectral ALS features in the prediction of diameter distributions. Instead, the findings showed that the features derived from bi-temporal ALS data (old leaf-off and recent leaf-on) achieved the lower mean of RMSE% values than the reference. The findings promote the knowledge how to apply remotely sensed datasets in the future forest inventories.