Species-related single dead tree detection using multi-temporal ALS data and CIR imagery

Abstract

The assessment of the health conditions of trees in forests is extremely important for biodiversity, forest management, global environment monitoring, and carbon dynamics. There is a vast amount of research using remote sensing (RS) techniques for the assessment of the current condition of a forest, but only a small number of these are concerned with detection and classification of dead trees. Among the available RS techniques, only the airborne laser scanner (ALS) enables dead tree detection at the single tree level with high accuracy.The main objective of the study was to identify spruce, pine and deciduous trees by alive or dead classifications. Three RS data sets including ALS (leaf-on and leaf-off) and color-infrared (CIR) imagery (leaf-on) were used for the study. We used intensity and structural variables from the ALS data and spectral information derived from aerial imagery for the classification procedure. Additionally, we tested the differences in the classification accuracy of all variants contained in the data integration. In the study, the random forest (RF) classifier was used. The study was carried out in the Polish part of the Białowieża Forest (BF).In general, we can state that all classifications, with different combinations of ALS features and CIR, resulted in high overall accuracy (OA ≥ 90%) and Kappa (κ > 0.86). For the best variant (CIR_ALSWSn-FH), the mean values of overall accuracy and Kappa were equal to 94.3% and 0.93, respectively. The leaf-on point cloud features alone produced the lowest accuracies (OA = 75–81% and κ = 0.68–0.76). Improvements of 0-0.04 in the Kappa coefficient and 0–3.1% in the overall classification accuracy were found after the point cloud normalization for all variants. Full-height point cloud features (F) produced lower accuracies than the results based on features calculated for half of the tree height point clouds (H) and combined FH.The importance of each of the predictors for different data sets for tree species classification provided by the RF algorithm was investigated. The lists of top features were the same, independent of intensity normalization. For the classification based on both of the point clouds (leaf–on and leaf-off), three structural features (a proportion of first returns for both half-height and full-height variants and the canopy relief ratio of points) and two intensity features from first returns and half-height variant (the coefficient of variation and skewness) were rated as the most important. In the classification based on the point cloud with CIR features, two image features were among the most important (the NDVI and mean value of reflectance in the green band).