191
Canopy Height Model
We generated a canopy height model (CHM) based on the overlapping single image
tiles. We used the software Postflight Terra 3D Vs4.0.104 (SenseFly 2015) to generate dense point clouds as *.las-files. Then we imported the *.las point cloud files
into the LAStools software (Isenburg 2016) to generate buffered tiles. Ground
points representing bare ground were identified visually. Based on these ground
points, we calculated the height (height normalization) for all non-ground points of
all tiles. These tiles were then mosaicked in SAGA-GIS using a b-spline interpolation with feathering to create a seamless normalized Digital Surface Model (nDSM).
This nDSM describes the maximum heights of the point cloud. Next, we generated
a Digital Terrain Model (DTM), that describes the minimum heights of the point
cloud. Finally, the CHM was generated by subtracting the DTM from the DSM,
which gave values in the range of −0.11 to 1.89 m. The lower range was adjusted to
zero. Average canopy height and its standard deviation were extracted for each canopy polygon.
Random Forest Classification
The Random Forest algorithm (Breiman Breimanx) is now a common standard nonparametric classifier with high performance as was found by many comparative
studies in a remote sensing context (Pal 2005; Duro et al. 2012; Qian et al. 2014).
Random Forest makes use of the concept of classification and regression trees
(CART) but combines them with ensemble modelling and bagging. Random Forest
is a non-parametric classifier that creates thousands of single decision trees and
averages their results. Each decision tree is a subsample of the whole dataset. The
split for each tree node is determined by the Gini criterion, which measures the
entropy of the dataset. The best split is that parameter value that leads to the largest
decrease in the Gini criterion. When the classifier is applied to the test dataset, the
final class label is then based on the majority vote of all constructed decision trees
(Immitzer et al. 2012).
A Random Forest classifier was used to predict species labels, with this achieved
by first dividing the dataset into training and testing polygons, with an 80:20 ratio
per class. To establish if any single set of parameters were sufficient alone, the dataset was split into a RGB, a NIR, a texture and a complete dataset (ALL). For quantifying the importance of the CHM, we added these values to each parameter dataset.
Before classification, all parameters with Pearson correlations higher than 0.75
were deleted to ensure that multicollinearity issues would not be an issue. Only two
texture parameters where omitted because of multicollinearity, the Cluster
Prominence and Haralick’s correlation. The latter was correlated with “correlation
(corrL)” and the first with “inverse distance moment (IDM)”. Finally, in order to test
the effect of species abundance on the classification results, the species data were
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species
Canopy Height Model
We generated a canopy height model (CHM) based on the overlapping single image
tiles. We used the software Postflight Terra 3D Vs4.0.104 (SenseFly 2015) to generate dense point clouds as *.las-files. Then we imported the *.las point cloud files
into the LAStools software (Isenburg 2016) to generate buffered tiles. Ground
points representing bare ground were identified visually. Based on these ground
points, we calculated the height (height normalization) for all non-ground points of
all tiles. These tiles were then mosaicked in SAGA-GIS using a b-spline interpolation with feathering to create a seamless normalized Digital Surface Model (nDSM).
This nDSM describes the maximum heights of the point cloud. Next, we generated
a Digital Terrain Model (DTM), that describes the minimum heights of the point
cloud. Finally, the CHM was generated by subtracting the DTM from the DSM,
which gave values in the range of −0.11 to 1.89 m. The lower range was adjusted to
zero. Average canopy height and its standard deviation were extracted for each canopy polygon.
Random Forest Classification
The Random Forest algorithm (Breiman Breimanx) is now a common standard nonparametric classifier with high performance as was found by many comparative
studies in a remote sensing context (Pal 2005; Duro et al. 2012; Qian et al. 2014).
Random Forest makes use of the concept of classification and regression trees
(CART) but combines them with ensemble modelling and bagging. Random Forest
is a non-parametric classifier that creates thousands of single decision trees and
averages their results. Each decision tree is a subsample of the whole dataset. The
split for each tree node is determined by the Gini criterion, which measures the
entropy of the dataset. The best split is that parameter value that leads to the largest
decrease in the Gini criterion. When the classifier is applied to the test dataset, the
final class label is then based on the majority vote of all constructed decision trees
(Immitzer et al. 2012).
A Random Forest classifier was used to predict species labels, with this achieved
by first dividing the dataset into training and testing polygons, with an 80:20 ratio
per class. To establish if any single set of parameters were sufficient alone, the dataset was split into a RGB, a NIR, a texture and a complete dataset (ALL). For quantifying the importance of the CHM, we added these values to each parameter dataset.
Before classification, all parameters with Pearson correlations higher than 0.75
were deleted to ensure that multicollinearity issues would not be an issue. Only two
texture parameters where omitted because of multicollinearity, the Cluster
Prominence and Haralick’s correlation. The latter was correlated with “correlation
(corrL)” and the first with “inverse distance moment (IDM)”. Finally, in order to test
the effect of species abundance on the classification results, the species data were
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species