AI-aided forest inventory with TerraScan: combining 3D and 2D features for tree species classification

In recent years, numerous methods have been developed to automate tree species classification using Shallow and Deep Learning techniques. Traditional approaches often utilize scanner-measured data (e.g., intensity) and 3D geometric features to compute statistical descriptors, which are used to train algorithms like Random Forests or Support Vector Machines. Deep Learning approaches, such as convolutional neural networks, process 2D raster images of point clouds but may lose critical 3D geometric details. Graph-based methods and architectures directly processing unstructured 3D data have shown promise but are often computationally intensive and less practical for industry.

To address these challenges, we developed a method that combines 2D raster and 3D point cloud features, achieving over 90% average classification accuracy. Our approach leverages well-established techniques and integrates them into TerraScan software for industrial use. Data augmentation, including SMOTE, addresses class imbalances, while features extracted from multiple raster viewpoints enhance dataset diversity.

Using TerraScan, users can efficiently preprocess data, augment training examples, and train models for over 10,000 trees in under 40 minutes on a GeForce RTX 4080. The system provides confidence scores with predictions, enabling manual evaluation of low-confidence results. This versatile method shows potential for broader object classification tasks beyond tree species identification.