Harvest cost surrogate
An XGBoost model trained to predict the output of FRCS (Forest Residue Cost Simulator) across the full parameter space of California cluster supply. The surrogate is what makes FRED capable of evaluating millions of clusters per query.
#What it replaces
FRCS is the USDA Forest Service tool for computing harvest cost given terrain, harvesting system, prescription, and biomass yield. It is deterministic, engineering-grounded, and considered the standard for forest biomass cost estimation in California.
FRCS is also slow. Each invocation requires loading the FRCS runtime, which is a multi-second operation. For a system that needs to evaluate procurement economics across millions of clusters per query, calling FRCS directly is not feasible.
FRED's harvest cost surrogate replaces FRCS-in-the-loop with a learned approximation that runs in roughly 0.2 milliseconds per prediction.
#Why a surrogate
The standard argument for surrogate modeling is speed. That holds here, but it is not the most important argument.
The more important argument is portability. The surrogate is a 12 MB serialized model file. It can run anywhere with an XGBoost binding: a Python notebook, a JavaScript backend, a research workstation, a serverless function. FRCS itself is a packaged tool with its own runtime requirements; running it at scale or in a non-standard environment is genuinely difficult.
The surrogate is a queryable, distributable artifact. FRED's California analysis runs through it, but researchers in other states can retrain it on their own FRCS runs without changing the rest of the pipeline.
#Training data
The training set is approximately 1.4 million FRCS configurations spanning the Cartesian product of:
- Terrain classes: flat, moderate, steep, very steep.
- Harvesting systems: cable, ground-based, helicopter, mixed.
- Prescriptions: every treatment ID in the C-BREC catalog.
- Biomass density per acre: sampled across the range observed in California clusters.
- Cluster size: sampled to cover small-block and large-block cases.
- Haul distance: sampled to cover near and far haul scenarios.
- Slope: sampled within each terrain class.
FRCS was invoked for every cell in this grid. Infeasible combinations (helicopter on flat ground, for instance) were retained as labeled-infeasible examples for the first-stage classifier.
#Two-stage architecture
The surrogate is two models in sequence:
- Feasibility classifier. Binary XGBoost classifier predicting whether a (terrain, harvest system, prescription) combination is operationally feasible. Trained on the full 1.4M configuration set with feasibility labels from FRCS. Held-out accuracy is 99.7 percent.
- Cost regressor. XGBoost regressor predicting per- BDT levelized cost for feasible configurations. Trained only on the feasible subset to avoid extrapolation errors.
The two-stage design matters because the cost surface is highly discontinuous across the feasibility boundary. A single model trying to predict cost across infeasible regions would have to either extrapolate wildly or compress the feasible distribution. Separating the two yields much better feasible-region accuracy.
#Validation results
Validation used a stratified 80/10/10 train/validation/test split with stratification on terrain class and harvesting system, so every terrain-system combination appears in the test set:
- R² (test, feasible configurations): 0.997
- Mean absolute error: under $2.10 per BDT
- P95 absolute error: under $6.50 per BDT
- Worst single-cluster error: $28.40 per BDT, on a steep-terrain helicopter operation at the edge of the biomass density distribution. Outliers like this exist; they are rare and the surrogate's confidence is appropriately low in those regions.
Additional held-out validation against fresh FRCS runs (not in training) shows no degradation. Full methodology and per-stratum error breakdowns are in Chapter 4 of the thesis.
#When to trust it
For typical California cluster configurations, the surrogate's predictions are dependable to within a few dollars per BDT. The regions of higher uncertainty are:
- Edge-of-distribution biomass densities. Very sparse or very dense clusters are less common in training and have wider prediction intervals.
- Marginally feasible configurations. Cases right on the feasibility boundary can have noisy cost predictions even though the classifier rates them as feasible.
- Helicopter operations on edge terrain. Helicopter cost in FRCS itself is more variable, and the surrogate inherits that variance.