Airborne dual-wavelength waveform LiDAR improves species classification accuracy of boreal broadleaved and coniferous trees

Abstract

Tree species identification constitutes a bottleneck in remote sensing applications. Waveform LiDAR has been shown to offer potential over discrete-return observations, and we assessed if the combination of two-wavelength waveform data can lead to further improvements. A total of 2532 trees representing seven living and dead conifer and deciduous species classes found in Hyytiälä forests in southern Finland were included in the experiments. LiDAR data was acquired by two single-wavelength sensors. The 1064-nm and 1550-nm data were radiometrically corrected to enable range-normalization using the radar equation. Pulses were traced through the canopy, and by applying 3D crown models, the return waveforms were assigned to individual trees. Crown models and a terrain model enabled a further split of the waveforms to strata representing the crown, understory and ground segments. Different geometric and radiometric waveform attributes were extracted per return pulse and aggregated to tree-level mean and standard deviation features. We analyzed the effect of tree size on the features, the correlation between features and the between-species differences of the waveform features. Feature importance for species classification was derived using F-test and the Random Forest algorithm. Classification tests showed significant improvement in overall accuracy (74→83% with 7 classes, 88→91% with 4 classes) when the 1064-nm and 1550-nm features were merged. Most features were not invariant to tree size, and the dependencies differed between species and LiDAR wavelength. The differences were likely driven by factors such as bark reflectance, height growth induced structural changes near the treetop as well as foliage density in old trees.