How can methods of species prediction be generalized and accurately transferred across core sites?
Species classification is a cornerstone in decision-making for environmental conservation as well as for many scientific, and management activities for forest managers. The aim of this paper was to develop a generalized classification model using airborne laser scanning (ALS) data to provide species or genus level tree identification. We attempted to develop a single generalized classification model that uses monospectral or multispectral ALS 3D and intensity features through a random forest algorithm to identify all individual tree species found in two study areas. Our analysis shows that a generalized method can identify nine species with a 73% overall accuracy, whereas the site-specific overall accuracies were 75% and 66% respectively.
The goal of this research was to explore standardization approaches for airborne laser scanning (ALS) feature values (“classification metrics”). Standardization approaches could help to reduce the field sampling required for such classification across different areas of interest (AOI). Earlier studies revealed that tree species were classified with moderate to high accuracy using AOI specific random forest models and ALS data. Classification accuracy varies from AOI to AOI, and model to model. However, field sampling and model development for each AOI still represent significant costs and time. This raises questions about ways to improve on this situation. First, do feature values for the same species differ between AOIs within a same ecological region, and is this variation different for 3D and intensity features? Second, can we train a classifier using samples of a given AOI and apply it to another? Third, can we train a classifier from a sample composed of trees from different AOIs (defined as a global sample)? To answer the above questions, we extracted both 3D and intensity features from multispectral and monospectral ALS data to identify trees. We standardized the feature values across three AOIs to minimize sampling needs and resolve classification issues. Three feature standardization approaches were employed: histogram matching, median-based adjustment, and regression-based adjustment (Figure 1). The improvements brought about by these methods were assessed through the Bhattacharyya distance (before vs. after standardization). The effects of standardization were also evaluated by analyzing the changes in the classification errors of random forest models (i.e. out-of-bag (OOB) error) for 11 tree species. The Bhattacharyya distance was reduced by between 17% to 27% for 3D features, and between 21% to 91% for intensity features. Training a random forest classifier using a sample from one AOI allowed OOB decrease of 14-21% when applied to other AOI after standardization. An OOB decrease after standardization of 4-20% was obtained when using a global sample. Our study reveals the prospect and challenges of different feature standardization methods for multispectral/monospectral ALS sensor to reduce field sampling at new AOIs while maintaining a good accuracy.
Figure 1. Feature standardization across areas of interest to optimize field sampling for individual tree species classification. YRF-York Regional Forest, PRF-Petawawa Research Forest. Bd-Tilia americana, Bw-Betula papyrifera, La-Larix laricina, Mh-Acer saccharum, Or-Quercus rubra, Pr-Pinus resinosa, Pt-Populus tremuloides, Pw-Pinus strobus, Sw-Picea glauca. I_NDG1_1st_mn – Green Normalized Difference vegetation index using first return.
List of publications
- Peer reviewed conference paper accepted in IGARSS 2018. Title: Towards a generalized method for tree species classification using multispectral airborne laser scanning in Ontario, Canada. https://igarss2018.org/
- Submitted an abstract to the ForestSAT 2018 conference. Abstract title: Feature standardization across areas of interest to optimize field sampling for individual tree species classification. http://forestsat2018.forestsat.com/