3D-MFDNN: Three-dimensional multi-feature descriptors combined deep neural network for vegetation segmentation from airborne laser scanning data

Airborne laser scanning (ALS) is a state-of-the-art technique for fast and accurate three-dimensional information acquisition of land cover including vegetation. This paper presents a three-dimensional multi-feature descriptors combined deep neural network-based methodology (3D-MFDNN) for ALS data-based vegetation segmentation with three well-designed steps namely generation of feature descriptors, 3D-MFDNN method's training, testing, and performance comparison using ALS datasets. The proposed 3D-MFDNN method is straightforward to implement, where accurate segmentation of tree points are effectively dealt in several complex cases, such as tree branches connected with other objects, tree with understory low vegetation, low-lying plants on the sloping surface, large tree with volumetric shape and branches hanging near sloping ground surface, etc. The method performance was evaluated using six datasets having different levels of scene complexity, and vegetation segmentation was performed at F1-score and accuracy of 83.94 % and 92.13 %, respectively. The method achieves significant improvement in comparison with several state-of-the-art methods.

Automated summary

This paper proposes a method called 3D-MFDNN for vegetation segmentation based on ALS data. The method generates feature descriptors, trains the 3D-MFDNN model, and performs testing and performance comparison using ALS datasets. It achieves accurate segmentation of tree points in complex cases and outperforms several state-of-the-art methods, with F1-score and accuracy of 83.94% and 92.13%, respectively, on six datasets with different levels of scene complexity.