Classification

Classification is the transformation between two states of the same ALS tile. Before, the tile is roughly a hundred million indistinguishable measurements in a coordinate system. After, it is a Digital Terrain Model with vegetation pulled away, building footprints isolated for extraction, and a powerline catenary tagged for utility inventory. Everything in this chapter is machinery for carrying the tile across that gap, and the cost of carrying it badly is concrete: classification errors propagate into every product derived from the tile, so a steep ridge labelled as building produces a phantom structure downstream, and canopy mistaken for ground produces a DTM that every contour drawn on it inherits.

The methods that perform the transformation span thirty years, from the hand-crafted-features-plus-Random-Forest pipeline that still ships in operational software and reaches above 80% overall accuracy on ALS benchmarks with limited training data (Weinmann et al., 2014), to the deep architectures that lead the Semantic3D benchmark of Hackel and colleagues (2017), where the strongest methods now exceed 76% mIoU (Thomas et al., 2019).

Classification sits near the end of the pipeline and draws on almost everything before it. It consumes the geometric features of Chapter 10, the height above ground that the terrain models of Chapter 9 provide, and, when labels are assigned per segment rather than per point, the segments of Chapter 11. Its output feeds object reconstruction (Chapter 16), change detection (Chapter 14), and the application workflows of Chapter 17, and the deep architectures it surveys in brief are developed fully in Chapter 18. The chapter builds from the bottom up: feature engineering first, then the training methodology that turns labels into a trustworthy model, the classifiers themselves (Random Forest and its gradient-boosted relatives, SVM, CRF regularisation), their deep competitors, and finally the accuracy assessment that defines the OA, mF1, and mIoU figures quoted along the way. We start with the features, because no classifier I have ever trained, however capable its architecture, has compensated for an uninformative input vector.

The classification pipeline

Traditional point cloud classification produces the label assignment defined above through a three-stage pipeline:

1. Feature extraction. For each point, compute a feature vector $\mathbf{f}_i \in \mathbb{R}^d$, the list of $d$ numbers that describes the local geometry, radiometry, and context and that the classifier sees in place of the raw point.
2. Supervised learning. Train a classifier (Random Forest, SVM, neural network) on a labelled training set $\{(\mathbf{f}_i, y_i)\}$ where $y_i$ is the class label. Supervised means the model learns from examples whose correct labels are already known, in contrast to unsupervised methods such as the clustering of Chapter 11, which receive no labels.
3. Spatial regularisation. Optionally smooth the per-point predictions using spatial context (CRF, majority voting), ensuring that neighbouring points tend to share the same label.