1 of 1

MCF = Thousand Cubic Feet

r = Mean Procurement Radius

INTRO

This analysis uses response-level data from the USDA Forest Service’s Timber Products Output Survey (2019-2021), namely self-reported mean procurement radius (miles) and annual procurement volume (MCF), to predict aggregate demand for non-pulp* roundwood at primary processing mills within the USDA Forest Service, FIA** Northern Region.

Outputs will aid policy-makers concerned with preserving heavily-harvested forested areas, as well as those within the industry who may seek to source logs closer to home, ideally minimizing haul distance.

METHODS

Given qualified responses account for just ~1/6 of the sample, procurement radius imputations for non-respondents is necessary.

Thus, regression analyses of varying complexity are utilized, with the leader chosen for sample-wide imputation.

All entities now house a measure of procurement radius, whether a direct response or model-based prediction.

Next, mill locations are geocoded, drawing a buffer around each mill corresponding to the imputed/response procurement radius. These buffers are further limited to only include areas from which a truck could arrive at the mill in one hour or less and are clipped to the outline of the US.

All areas of radius-coverage are weighted by the SqRt of the mill’s annual procurement volume, SqRt(MCF). All weighted areas of radius-coverage are then summed to produce a final demand ‘heat map’.

RESULTS

Demand for non-pulp* roundwood varies substantially at a geographic scale. Notably, key ‘wood baskets’ exist in the region of coverage.

DISCUSSION

Given demand for non-pulp* roundwood is concentrated in prominent ‘wood baskets’, relocation of select primary processing mills and/or log suppliers may be warranted.

Quantifying Aggregate Demand for Non-Pulp* Roundwood at Primary Processing Mills in the Northern US – A Machine Learning Approach

Models Assessed

4. Random Forest Reg. (5x CV):

sqrt(r) ~ x

          • Tuned # of trees
          • Tuned ‘mtry’

    • Gradient Boosted Reg. (5x CV):

sqrt(r) ~ x

    • Machine Learning (h2o) Reg. (5x CV):

sqrt(r) ~ x

Preliminary Results

 

Ian Kennedy, MS Geography

Research Assistant, FFRC

iankennedy@umass.edu

i = Region (NE, Mid Atl, Central, Lake, Plains)

x = log10(MCF), State, Mill Type (Sawmill/Other), Lat/Lon, # of employees, Equipment, Portable?, Logs exported?

Scan for project overview^

Sample Statistics (Procurement Radius)

Sample Size - 2,586

Qual. Respondents - 427

Nonrespondents - 2,159

Normal QQ-Plots:

  1. Sample-wide procurement volume
    • Log10(MCF)
  2. Qual. Respondents’ procurement radius:
    • SqRt(Pro. Radius)

All regression analyses were conducted in RStudio. Mapping was completed in ArcGIS Pro, through usage of a Python-based script.

*Non-Pulp = saw, house/cabin, post, pole, & res. firewood logs*

**FIA = Forest Inventory & Analysis**