Adding single tree features and correcting edge tree effects enhance the characterization of seedling stands with single-photon airborne laser scanning

A New Approach for Feeding Multispectral Imagery into Convolutional Neural Networks Improved Classification of Seedlings

“…This was because the C th -affected tensors of this study comprised only 33.4% of the dataset, and the rest of the data were not affected as their cells were all above the C th of 0.4 m. It should be noted that apart from one of the forest plots in this study (1 of 75 plots), all of the plots were in advanced seedling (mean tree height > 1.3 m, AdS) stands, while in the study by Imangholiloo et al [2], there were four plots in young seedling (mean tree height ≤ 1.3 m, YoS) stands plus ten plots in AdS. Hence, the methodology developed in this study could help in inventorying YoS more accurately, although it was documented to be challenging for characterizing seedling stands using hyperspectral and RGB photogrammetric point clouds [46] and using ALS [2,45].…”

“…The applied methodological improvement in this study-using canopy threshold (C th )-based image pre-processing prior to feeding to CNN classifiers-was inspired by other seedling studies using ALS data. The studies usually applied C th s between 0 and 1 m for the removal of laser returns from understory/ground [2,10,[42][43][44][45]. C th s between 0.5 and 1 m were also applied in seedling classification using conventional classifiers such as RF on hyper-or multi-spectral data [46].…”