Effect of feature standardization on reducing the requirements of field samples for individual tree species classification using ALS data

The objective of this study was to evaluate the effectiveness of three standardization approaches for airborne laser scanning (ALS) feature values used for individual tree species classification. This study is the first effort to assess the transferability of forest tree species classification models derived using monospectral and multispectral ALS data. Three research questions were asked; (1) How do the ALS features differ for the same species in different though comparable ecological regions? (2) How to train a model with one sub-population and apply it in another sub-population? (3) How to fuse models for two areas into a global model? To answer these questions, both 3D and intensity features were extracted from the ALS data from Canadian boreal forests. The ALS feature values were standardized in two different scenarios, disjoint areas, and partially overlapping areas, across three study areas. Feature standardization approaches were used: histogram matching, median-based standardization, and linear regression-based standardization. A linear discriminant analysis (LDA) and random forest (RF) algorithms were employed to classify the study area’s major tree species. The Bhattacharyya distance and overall accuracy (OA) were used to assess the classification model performance before and after the feature standardization. Three major conclusions were drawn. First, the Bhattacharyya distance confirmed that intensity features varied across study areas and among tree species, while 3D features were relatively less variable. Second, for the disjoint areas (York Regional Forest (YRF)) and Petawawa Research Forest (PRF)), the feature standardization procedure consistently improved the OA classification for both local model and global model approaches. The feature standardization improved the OA from 16% to 54% using LDA, and from 20% to 55% using RF in the local model. The improvement was from 58% to 66% using LDA, and from 63% to 70% using RF in the global model. It can be concluded that intensity features (at YRF and PRF) were most prone to differ between areas because of scanners and acquisition settings. If ALS data were available from both areas, intensity features need to be normalized so that the local model can be transferred. Finally, for the partially overlapping areas (the northern and southern parts of Black Brook Forest), this study suggests that normalization of ALS data is not needed because they were captured using quite similar ALS settings.

2022