Datasets:

taylor-geospatial
/

coordbench

To evaluate performance on biodiversity-related applications, we leveraged the Global Biodiversity Information Facility (GBIF) records, specifically research-grade observations from the iNaturalist citizen science repository (1). GBIF is a comprehensive collection of species occurrence records. We focus on genus-level taxonomic labels in the United States given interest in forest species composition mapping across a diversity of forest types.

We select GBIF records for the period Jan 1 2017 to Jan 1 2023 and filter to just observations where the genus label is found in a list of tree genera sourced from the US Forest Service and country code is set to the US. Observations must be labeled as human or machine observations, and have a maximum spatial uncertainty of 10 meters. From this initial list of tree genera observations, we get the five most frequently observed species for each of the US states, including Alaska and Hawaii, then combine into a single deduplicated list of common US tree genera. We select these genera for further processing.

Of the remaining observations across common tree genera, we drop any that have less than 500 samples per class, resulting in a final set of 39 genera. We allocate 300 samples to the train split and the rest to the test split. We treat observations as instantaneous labels, using the date of the observation record ("eventdate") as a single-date valid period. Our final us_trees evaluation dataset has a total of 11,700 training points and 33,682 test points after pre-processing and spatial proximity filtering.

Table: us_trees classes and sample counts by split

label	label_name	train_count	test_count
0	abies	300	827
1	acer	300	1296
2	aesculus	300	665
3	ailanthus	300	857
4	alnus	300	220
5	amelanchier	300	203
6	asimina	300	614
7	betula	300	917
8	carya	300	628
9	cercis	300	1296
10	cornus	300	1296
11	diospyros	300	934
12	elaeagnus	300	1296
13	fagus	300	1187
14	gleditsia	300	353
15	ilex	300	1296
16	juglans	300	620
17	juniperus	300	1296
18	liquidambar	300	1296
19	liriodendron	300	1173
20	maclura	300	317
21	magnolia	300	874
22	morus	300	387
23	picea	300	621
24	pinus	300	1296
25	populus	300	1296
26	prosopis	300	1057
27	prunus	300	1296
28	pseudotsuga	300	602
29	quercus	300	1296
30	sabal	300	272
31	salix	300	957
32	sassafras	300	1006
33	taxodium	300	271
34	thuja	300	474
35	triadica	300	300
36	tsuga	300	1190
37	ulmus	300	604
38	yucca	300	1296

License

iNaturalist is licensed under a Creative Commons Attribution Non-Commercial 4.0 License (CC-BY-NC). You may obtain a copy of the CC-BY-NC license at https://creativecommons.org/licenses/by-nc/4.0/deed.en. You can obtain a copy of the dataset at: https://www.gbif.org/dataset/50c9509d-22c7-4a22-a47d-8c48425ef4a7#description. This version of the dataset has been modified as described above.

For the dataset citation, please see the “References” section below.

References

iNaturalist contributors, iNaturalist Research-grade Observations, iNaturalist.org (2025); https://doi.org/10.15468/ab3s5x