We recently conducted a Trimble eCognition webinar called Timber Cruising with eCognition where I explored the use of remote sensing data to automatically extract valuable information on the individual tree level. A combination of LiDAR and imagery data was used to extract valuable attributes for each tree such as height, position on the ground and crown size.
During our webinar, a number of attendees asked about the potential of classifying tree species at the tree level. To add this level of detail to the analysis depends greatly on the input data. A number of studies have explored the use of hyperspectral data. One such example was reviewed in our blog post “Tree Species Classification with Trimble eCognition”. But, the use of hyperspectral data can be prohibitive as it is not widely available and/or fully integrated to the mainstream of remote sensing.
Shortly after the webinar, I was contacted by one of our customers in Quebec, Canada – Centre D’enseignement et de Recherche en Foresterie de Sainte-Foy (CERFO). Mathieu Varin and his colleagues from CERFO recently published a paper in the MDPI remote sensing journal entitled “Object-Based Approach Using Very High Spatial Resolution 16-Band WorldView-3 and LiDAR Data for Tree Species Classification in a Broadleaf Forest in Quebec, Canada” where they have developed a method to identify 11 tree species based on rich multispectral data in combination with LiDAR.
The authors point out that although photo-interpretation has been used for forest characterization since the last century, the primarily manual method has unfortunately not changed much and tree species identification is still an open issue. In this paper, they investigate the unique data fusion capabilities of eCognition to combine 16-band WorldView-3 imagery with LiDAR point clouds to evaluate improvements to single tree classification, particularly as it concerns tree species identification. “The classification of species is divided into three parts, on two levels: (1) tree types and (2) broadleaf and conifer tree species”. The study focuses on the development of a machine learning based approach, looking at various supervised classification techniques.
The initial part of the research examines the optimal segmentation inputs to attain single tree objects and shows that the best results are achieved when combining both the Canopy Height Model (CHM) and multispectral data in the segmentation algorithm. With the Individual Tree Crown (ITC) segmentation based on the corrected CHM and image data “82% of the objects represented a single species”.
From here, various supervised classification models were explored. The authors determined that the Random Forest (RF) model yielded the highest accuracy in both the global and hierarchical approaches:
- “For the global modeling approach, RF (ntree: 2000; mtry: 3) was selected as the best model based on the performance assessment, as it gives an OA of 75% and a KIA of 0.72 when using 16-band derived variables”.
- “For the hierarchical modeling approach, RF (ntree: 2000; mtry: 2) presented the best performance with OA of 99% and KIA of 0.97 for tree type (broadleaf/conifer) modeling”.
In conclusion, the authors believe that “this method could also be applied on a large scale with limited manipulations. The resulting maps represent a valuable tool with which to analyze forest composition and to guide forest planners”.
I would like to thank Mathieu and his team for sharing their research as it relates to the discussion we recently had in our webinar. It proves that tree level species classification is possible with rich multispectral remote sensing data and is improved with the integration of LiDAR. I encourage you to read their full paper for all the details on this very interesting research.