Modelling Natural History to Increase Native Seed Collection Efficiency - Space, Time, and Connectivity

Natural Areas Conference

Reed Clark Benkendorf

Senior Spatial Data Specialist & Senior Botanist

October 8, 2024

Background

• Remote Sensing, Citizen Science, Statistics, and Ecology
  
• research projects focus on theory and shoot for the moon
  
• West very large, lacking distribution data on most species
  
  
• What’s needed for seed collection?
  

• Better knowledge of Natural History

Lot’s of absolutely amazing things are happening in ecology these days. From the proliferation of Citizen Science initiatives to fancy fancy remote sensing, and amazing advances in statistics. However, I generally feel that the developers of these things get carried away with testing these out on theoretical problems, whereas we have many real problems. In particular, we’ve still much work to do in documenting basic natural history - a feat largely abandoned by the main stream 40 years ago.

Most all of my work is wrangling museum, government, and citizen science data to try and share natural history information with seed collectors - in quick, easy to access digital formats.

SPACE

“It is advisable to look from the tide pool to the stars and then back to the tide pool again.”

John Steinbeck & Ed Ricketts

Challenge

• Want: 20 collections per species per seed transfer zone
• Native species declining - wildfires, habitat degradation
• Many populations needed as drought, pests, etc. vary each year
• Occurrence data often biased against BLM

Johanna Steensma

On most BLM administered land, we simply have difficulty finding enough populations to collect from. Even populations of once wide-ranging abundant species can be difficult to find in the right condition for a crew to collect from. SOS official guidance is that crews can hope to make collections from 1/4 populations which are large enough, the other 3 are often lost to pests or other eccentricities of the year.

While some good occurrence data exist, once you really zoom in - you realize it’s pretty sparse. For these reasons we need to get crews lot’s of good ideas on where to go scout for their target species.

Solution

Model Call

\[\text{Species (Pres.|Abs.) ~}\\ \text{geology * topography * soils * vegetation * climate}\]

• Species Distribution Modelling (SDM’s), ENM’s etc.
• Predict the probability of suitable habitat
  
• Modelled 353 taxa at 90m resolution
  

Workflow

Independent Variables

Description	Source
Mean Annual Air Temp. (BIO1)	Chelsa
Temp. seasonality (BIO4)	Chelsa
Max Temp. of Warmest Month (BIO5)	Chelsa
Min Temp. of Coldest Month (BIO6)	Chelsa
Mean Temp. of Warmest Quarter (BIO10)	Chelsa
Mean Temp. of Coldest Quarter (BIO11)	Chelsa
Mean annual precip. (BIO12)	Chelsa
Precip. of Warmest Quarter (BIO18)	Chelsa
Precip. of Coldest Quarter (BIO19)	Chelsa
Mean Monthly vapour pressure deficit (vpd)	Chelsa
Heat accumulation of degree-days > 5C (gdd5)	Chelsa
First growing degree day > 5C (gdgfgd5)	Chelsa
Number of degree-days > 5C (ngd5)	Chelsa
Heat accumulation of degree-days > 10C (gdd10)	Chelsa
First growing degree day > 10C (gdgfgd10)	Chelsa
Number of Degree-days > 10C (ngd10)	Chelsa
Mean monthly near surface humidity (hurs)	Chelsa
Number of Days with Snow Cover(scd)	Chelsa
Annual Snow Water Equivalent (swe)	Chelsa
Percent Herbaceous Vegetation	EarthEnv
Percent Shrub Cover	EarthEnv
Percent Tree Cover	EarthEnv
Soil Depth to Bedrock (R Horizon)	SoilGrids
Soil organic carbon (Tonnes / ha)	SoilGrids
Soil Surface (0-5 cm) pH in H2O	SoilGrids
Soil 30-60 cm pH in H2O	SoilGrids
Soil Surface (0-5 cm) % clay	SoilGrids
Soil 5-15 cm % clay	SoilGrids
Soil Surface (0-5 cm) % silt	SoilGrids
Soil 5-15 cm % silt	SoilGrids
Soil 15-30 cm % silt	SoilGrids
Soil Surface (0-5 cm) % sand	SoilGrids
Soil 5-15 cm % sand	SoilGrids
Soil 15-30 cm % sand	SoilGrids
Soil Surface (0-5 cm) coarse fragments	SoilGrids
Soil (30-60 cm) coarse fragments	SoilGrids
Soil Salinity	SoilGrids
Elevation	MERIT - DEM
Slope	Geomorpho90
Aspect	Geomorpho90
Topographic Position Index	Geomorpho90
Compound Topographic Index	terra
Topographic Roughness Index	terra
Human Influence Index	NASA Earth Data

Dependent Variables

Source	No. Records
Herbaria	245,048
iNaturalist	171,981
AIM (BLM)	195,544
VegBank	55,124
FIA (USFS)	1,521

Our solution is to communicate the plausible range of individual species to our young collectors. We accomplish this by using Species Distribution Models - often also called Environmental Niche Models, which can be readily interpreted by crew leads using a GIS.

The SDMs we will discuss are constructed from a subset of generally 10-20 predictor variables selected from a raster stack of 44. We model these SDM’s using random forests, which are a very very fast algorithm, and provide reproducible results. The training data come from a variety of musuem and government sources.

Implementation

• Crew Leads receive output raster data (two formats)
  • Masked to >70% prob. of suitable habitat
  • Raw probability predictions
    
• Also receive raw training presences
  
• Visually review data while preparing hitches
  

Results

TIME

“The two most powerful warriors are time and patience”
Leo Tolstoy

Challenge

• Crews spend 15-25% of work driving
  
• Leads often overwhelmed with number (gen. 40-80) of target species
  
• Finite windows for scouting and collecting
  
• Fruit develops simultaneously
  

Katie Peel

So time is really of the essence for our crews. To be blunt, most of our seed matures and disperses in a couple batches, once in June to early July, and again in September - if you are in a monsoonal area. It is imperative that crews have found many populations before this happens, and hence that they prioritze their scouting efforts. Searching for the earliest species as soon as possible, and holding off on the later material.

Solution

Model Call

\[\text{GAM(Flowering (0|1) ~} \text{ DOY *}\\ \text{GDD * VPD * BIO10 * BIO14 * Soil Bulk Density * CTI * } \\ \text{corr. = corARMA(~ 1|year) * corExp(~ Lat.|Long.))} \]

• Generalized Additive Models
• describe the onset, shape, and cessation of a phenophase via wiggles
• Modelled ca. 270 species at 250m resolution

Workflow

Independent Variables

Description	Source
Mean Temp. of Warmest Quarter (BIO10)	Chelsa
Precipitation of driest month (BIO14)	Chelsa
Mean monthly vapour pressure deficit (vpd)	Chelsa
Heat accumulation of degree-days > 5C (gdd5)	Chelsa
First growing degree day > 5C (gdgfgd5)	Chelsa
Number of degree-days > 5C (ngd5)	Chelsa
Heat accumulation of degree-days > 10C (gdd10)	Chelsa
First growing degree day > 10C (gdgfgd10)	Chelsa
Number of Degree-days > 10C (ngd10)	Chelsa
Soil Bulk Density	SoilGrids
Compound Topographic Index	terra

Dependent Variables

Source	No. Records
Herbaria	64,529

This year the only phenophase we modelled was flowering, but an identical approach is suitable for fruiting or other phases.

Implementation

• Internally, GAMs are predicted into space
• These are subset to initiation and end dates
• Values are extracted from relevant layers for a spreadsheet

Results

• “all models are wrong, some are useful” how do you want to be wrong?

• Few absence records collected; highly un-balanced data set

CONNECTIVITY

“In nature we never see anything isolated, but everything in connection with something else which is before it, beside it, under it and over it.”

Johann Wolfgang von Goethe

Challenge

• Species Distribution Models only identify areas with theoretically suitable habitat
• Do not account for migration
  
• Modeled habitat != occupied habitat

Johanna Steensma

A species distribution model produces one and one metric only - the probability of suitable habitat BASED upon the dependent observations and independent variables you feed in. It suffers from both bias, especially due to where training records are located, and variance - associated with noise on the input data and over fitting of parameters. Our crews are focused on finding populations for collections - not mapping new populations ala a conservation focus. We want to get them there.

Notably then, we want crews to go where they have very high chances of being successful. The discrepancy between the name “species DISTRIBUTION model” and what’s it actually offers have been noted for roughly 20 years ago. In my experience most critiques of SDM’s arise from the mismatch between the nomenclature, and users expectations - and an SDM predicting suitable habitat very far from any currently occupied habitat.

Solution

• Reprocess the output of SDM’s to identify clusters of suitable habitat
  
• Identify populated patches
  
• Rank the connections and classify distance from populated to un-populated patches
  

Scoring Schema

Ranks for Contiguous Patches

Connection	No. Connections	Rank
1st	>= 2	2
1st	1	3
2nd	>= 3	4
2nd	<= 2	5
3rd	>= 4	6
3rd	<= 3	7

Ranks for non-contiguous patches,with occupied patches < 5km away

No. Occupied Patches	Rank
>= 3	8
<= 2	9

Workflow

Implementation

• Crew Leads receive output vector data of patch ranks
  • higher priority patches darker colors
    
• Visually review data while preparing hitches
  

Results

• Current hydrologic basins data set is to coarse
  • 400% more pops found in R1 than expected by the number of them
  • 74% more pops found R1 then expected by area alone
    

ACKNOWLEDGEMENTS

The Bureau of Land Management

Colorado Plateau Native Plant Program
Great Basin Native Plant Program
California Native Plant Program
Mojave Native Plant Program


Chris Woolridge, Kristy Snyder, Peggy Olwell

CONTACT