Treating a tree as a graph for biomass estimation

Estimating aboveground biomass for individual trees from LiDAR is a deceptively hard problem. The signal is there in the point cloud, but the structure that carries most of the meaning, the way a trunk splits into branches, is exactly what a flat point-based model has to rediscover from scratch every time.

The baseline, honestly

On the dataset I am working with at UPV, a plain random forest over handcrafted point-cloud features is a strong baseline. A point-based deep network did not automatically beat it. That result is worth sitting with rather than explaining away. If a simpler model wins, the deep model is not yet using the right inductive bias.

The graph view

A tree has a natural skeleton. Quantitative structure models give you that skeleton as a set of connected cylinders: where the trunk is, where it branches, how thick each segment is. That is already a graph. Nodes are segments, edges are the parent and child connections, and node features can carry length, radius, and orientation.

Phrased that way, the problem fits a graph neural network much better than a raw point cloud does. The model does not have to learn that a branch is attached to a trunk; the edge says so. It can spend its capacity on the harder question of how structure maps to mass.

What I am watching

A few things decide whether this is worth it:

• Whether the skeleton extraction is stable enough across trees of different species and sizes, because a noisy graph is worse than no graph.
• Whether a second dataset with different sensors and forest types confirms the approach instead of overfitting to one source.
• Whether the gain over the random forest is large enough to justify the extra pipeline complexity in the field.

None of that is settled yet. When it is, the full method and a reproducible repository will go into a proper paper. This is just the working view from the middle of it.