Laurence A. Clarfeld Katherina D Gieder Angela Fuller Jeff A Tracey Therese M Donovan 20250317 spreadsheet https://doi.org/10.5066/P1E7NDAF Laurence A. Clarfeld Jeff A Tracey Katherina D Gieder Angela Fuller Therese M Donovan 2025 application/service DeepFaune New England (DFNE) is a model for species classification in trail camera imagery. This model is a re-trained version of the DeepFaune model for classifying European species in trial cameras, fine-tuned to classify taxa from northeastern North America. The DFNE model takes as input cropped images of each animal, which can be generated by an object detection model. DFNE classifies 24 taxa, including the "no-species" label indicating the absence of an animal. This data release includes metadata for the training dataset used to train DFNE (.csv) and model parameters for running it (.pth). For more information on the DeepFaune New England model, see the software release available at: https://code.usgs.gov/lclarfeld/deepfaune-new-england-test This regional model fills a gap in existing openly licensed classification models for trail camera imagery and the data archived here provide reference for the training dataset metadata and model parameters. 20031113 20241026 ground condition Complete None planned Northeast North America -79.5410 -64.2480 48.0487 40.8471 ISO 19115 Topic Category biota None wildlife monitoring camera trap trail camera image classification machine learning species distribution modeling USGS Metadata Identifier USGS:67ae17fcd34e3f09c0e0f002 None Vermont New York New Hampshire Massachusetts Maine None American Marten Birds American Black Bear Bobcat Coyote Domestic Cat domestic cattle (feral) domestic dog Fisher Gray Fox gray squirrel Human Moose Rodentia Virginia Opossum Raccoon Red Fox Red Squirrel Striped Skunk Snowshoe Hare White-tailed Deer wild boar Wild Turkey Kingdom Animalia Subkingdom Bilateria Infrakingdom Deuterostomia Phylum Chordata Subphylum Vertebrata Infraphylum Gnathostomata Superclass Tetrapoda Class Mammalia Subclass Theria Infraclass Eutheria Order Carnivora Suborder Caniformia Family Mustelidae Subfamily Guloninae Genus Martes Species Martes americana TSN: 180559 Genus Pekania Species Pekania pennanti TSN: 1086061 Family Ursidae Genus Ursus Species Ursus americanus TSN: 180544 Family Canidae Genus Canis Species Canis latrans TSN: 180599 Species Canis lupus Subspecies Canis lupus familiaris TSN: 726821 Genus Urocyon Species Urocyon cinereoargenteus TSN: 180609 Genus Vulpes Species Vulpes vulpes TSN: 180604 Family Procyonidae Genus Procyon Species Procyon lotor TSN: 180575 Family Mephitidae Genus Mephitis Species Mephitis mephitis TSN: 180562 Suborder Feliformia Family Felidae Subfamily Felinae Genus Lynx Species Lynx rufus TSN: 180582 Genus Felis Species Felis catus TSN: 183798 Order Artiodactyla Family Bovidae Subfamily Bovinae Genus Bos Species Bos taurus TSN: 183838 Family Cervidae Subfamily Capreolinae Genus Alces Species Alces americanus TSN: 898198 Genus Odocoileus Species Odocoileus virginianus TSN: 180699 Family Suidae Subfamily Suinae Tribe Suini Genus Sus Species Sus scrofa TSN: 180722 Order Rodentia TSN: 180130 Suborder Sciuromorpha Family Sciuridae Subfamily Sciurinae Tribe Sciurini Genus Sciurus Subgenus Sciurus (Sciurus) Species Sciurus carolinensis TSN: 180175 Genus Tamiasciurus Species Tamiasciurus hudsonicus TSN: 180166 Order Primates Suborder Haplorrhini Infraorder Simiiformes Superfamily Hominoidea Family Hominidae Subfamily Homininae Genus Homo Species Homo sapiens TSN: 180092 Order Lagomorpha Family Leporidae Genus Lepus Species Lepus americanus TSN: 180112 Infraclass Metatheria Order Didelphimorphia Family Didelphidae Subfamily Didelphinae Genus Didelphis Species Didelphis virginiana TSN: 179921 Class Aves TSN: 174371 Order Galliformes Family Phasianidae Genus Meleagris Species Meleagris gallopavo TSN: 176136 None. Please see 'Distribution Info' for details. None. Users are advised to read the dataset's metadata thoroughly to understand appropriate use and data limitations. Laurence A. Clarfeld Vermont Cooperative Fish and Wildlife Research Unit Post-doctoral Fellow mailing and physical

University of Vermont

Burlington VT 05405 802-656-3011 lclarfel@uvm.edu All model training was performed on the USGS Tallgrass Supercomputer, part of USGS Advanced Research Computing. The Vermont Cooperative Fish and Wildlife Research Unit is jointly supported by the U.S. Geological Survey, University of Vermont, Vermont Department of Fish and Wildlife, US Fish and Wildlife Service, and Wildlife Management Institute. The 1.1 GB model parameters file (dfne_weights_v1_0.pth) was produced with Python 3.10 using the PyTorch package (v 2.6.0) function "torch.save". Metadata come from numerous openly licensed projects with varied quality control measures and contain automatically-generated bounding boxes that may be incorrect. To minimize incorrectly labelled data, we took the step of omitting machine-generated labels with confidence scores below 0.75 (0-to-1 scale) and include confidence scores with the metadata which give some indication of accuracy. The metadata represent unique image files and the values of all fields fall within expected ranges. The files contain no duplicate records. The contained metadata are complete and represent all image files used to train the DeepFaune New England model. DATASET SELECTION AND ASSIMILATION Training data for the DFNE originated from 4 repositories and comprised 15 distinct datasets. Metadata for all datsets were converted to SQLite databases compatible with the R package AMMonitor (v 2.1). Each data repository, with specific processing steps, will be discussed. 1. AMMonitor ScienceBase Community Ten datasets originate from the "ScienceBase" repository and are part of the "AMMonitor Community". All images in this repository had been analyzed with MegaDetector (v 5a). These datasets are openly licensed (with the exception of images containing humans) and can be accessed from: https://www.sciencebase.gov/catalog/item/6188c0c4d34ec04fc9c4f7a4. The AMMonitor package includes the function "sbRehydrate" that can be used to create an AMMonitor database for each dataset directly from ScienceBase. 2. Labeled Information Library of Alexandria: Biology and Conservation (LILA) The LILA repository hosts numerous conservation and biology datasets, mostly containing labelled images, intended for use by conservation and ML researchers. The three trail camera datasets we included from LILA had already been analyzed with MegaDetector (v 5a). Metadata from the LILA repositories were used to populate an AMMonitor database. All LILA projects are openly licensed and available at: https://lila.science/ 3. iNaturalist The citizen science platform, iNaturalist, allows users to share biodiversity information, including photos of wildlife. We used the iNaturalist API to query photos taken between January 1, 2000 and September 1, 2023 that were (a) added to the project "Camera Traps (Trail-cams)"; (b) licensed [(CC-BY-NC)](https://creativecommons.org/licenses/by-nc/4.0/) allowing use; (c) were labelled as containing one of the target species; and (d) taken in North America. Our search yielded 14385 candidate observations that met our criteria and contained properly formatted images. File names for images in this project are the iNaturalist "photo_id". All images were analyzed by MegaDetector (v 5a). iNaturalist data is available from: https://www.inaturalist.org 4. Cooperative Fish & Wildlife Research Unit - New York (CRUNY) Data from over 600 sites in New York state were made available through a data sharing agreement. All images were analyzed by MegaDetector (v 5a). Metadata from the CRUNY repository were used to populate an AMMonitor database. These data were made available through special arrangement and are not publicly available. 20250101 DATA FILTERING AND SPLITTING Candidate training samples were generated by cropping the original trail camera image using a bounding box produced by MegaDetector. We discarded detections where either (1) the detection confidence score was below 0.75 (for images with animals); (2) the image did not contain an animal from our list of target taxa; or (3) the image contained more than one species of animal. We included false positives detections from MegaDetector (those predicted by to contain an animal/human that do not), regardless of confidence score. We then used importance sampling to randomly select a subset of images, with no more than 12.000 images per taxon, into training (80%), validation (10%), and testing (10%) sets that ensure the image locations were mutually exclusive between the sets while maintaining the same relative frequency of taxa within each set 20250101 MODEL INITIALIZATION We used the DINOv2 ViT model architecture as our base model and loaded the weights from the DeepFaune (v 1.2) model. We then randomly sampled 20 images of each taxon and analyzed them with the DeepFaune (v 1.2) model. If at least half (10/20) of the taxon's images were mapped to the same European taxon, the weights of the fully connected layer corresponding to that taxon's node were initialized with the weights from the coresponding node from the DeepFaune model. If the taxon was not successfully mapped to a European taxon, the weights for the corresponding neruron in the new model were initialized randomly. All weights were then perturbed with a random jitter sampled from a normal distribution with mean value equal to 1% of the initial weights. 20250101 MODEL TRAINING The fully-connected layer of the initialized model was re-trained, using the training set, with PyTorch (v 2.6.0) / Python (v 3.10). All training was performed on the Tallgrass HPC cluster. We used a stochastic gradient descent optimizer with a learning rate of 1e-4. After 25 training epochs, the model weights for the epoch with highest validation accuracy was selected as the final model. 20250101 dfne_training_dataset.csv Comma Separated Value (CSV) file containing data. Producer Defined project The dataset abbreviation (project) of the originating dataset Producer Defined caltech Caltech Camera Traps https://lila.science/datasets/caltech-camera-traps Beery, S., Van Horn, G. & Perona, P. (2018). Recognition in terra incognita. Proceedings of the european conference on computer vision (ECCV), pp. 456–473. Producer defined cruny Cooperative Fish and Wildlife Research Unit - New York Producer defined dart Dartmouth College Woodlands Wildlife Monitoring Project https://www.sciencebase.gov/catalog/item/663cde7ad34e77890839e157 Patry, R.K., Sirén, A.P.K., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, Dartmouth College Woodlands Wildlife Monitoring Project Volume 1 (2014 - 2024): U.S. Geological Survey data release, https://doi.org/10.5066/P133GQPK. Producer defined ENA24 ENA24-detection https://lila.science/datasets/ena24detection Yousif H, Kays R, Zhihai H. Dynamic Programming Selection of Object Proposals for Sequence-Level Animal Species Classification in the Wild. IEEE Transactions on Circuits and Systems for Video Technology, 2019. Producer defined gmnf USDA Green Mountain National Forest https://www.sciencebase.gov/catalog/item/663cdf96d34e77890839e178 Gieder, K.D., Bernier, C.A., Staats, S.A., Wixsom, S.J., Abrams, R.J., Cahill, J.R., Crumley, K., Royar, K., Siren, A.P.K., Wilson, T.L., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, USDA Green Mountain National Forest Volume 1 (2016 - 2022): U.S. Geological Survey data release, http://doi.org/10.5066/P1GVIBFL. Producer defined inat iNaturalist https://www.inaturalist.org/projects/camera-traps-trail-cams iNaturalist. (2024). iNaturalist observations dataset. Producer defined mass Massachusetts Wildlife Monitoring Project https://www.sciencebase.gov/catalog/item/6672de8dd34e84915adbb4f3 Wilson, T.L., Siren, A.P.K., Berube, J.A., Morrow, C.M., Wattles, D.W., Huguenin, M., Clarfeld, L.A., Huber, K.E., Donovan, T.M., 2024, Massachusetts Wildlife Monitoring Project (2022 - 2024): U.S. Geological Survey data release, https://doi.org/10.5066/P13UNTFB. Producer defined mdifw Maine Department of Inland Fisheries and Wildlife Moose Project https://www.sciencebase.gov/catalog/item/6669b08cd34e9bcc607bd881 Kantar, L.E., Sirén, A.P.K., Wilson, T.L., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, Maine Department of Inland Fisheries and Wildlife Moose Project - Volume 2 (2021 - 2024): U.S. Geological Survey data release, https://doi.org/10.5066/P132SU4S. Producer defined mefur Maine Department of Inland Fisheries and Wildlife https://www.sciencebase.gov/catalog/item/654e4ae5d34ee4b6e05c315a Webb, S.M., Clarfeld, L.A., Huber, K.E., and Donovan. T.M., 2024, Maine Department of Inland Fisheries and Wildlife Volume 1 (2022 - 2023): U.S. Geological Survey data release, https://doi.org/10.5066/P9ML66K3. Producer defined nhfg New Hampshire Fish and Game Department https://www.sciencebase.gov/catalog/item/6670b566d34e89718fa3fa7a Jones, H., Sirén, A.P.K., Callahan, C.B., Holman, H., Marchand, M.N., Kilborn, J.R., Wilson, T.L., Morelli, T.L., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, New Hampshire Fish and Game Department Volume 1 (2014 - 2024): U.S. Geological Survey data release, https://doi.org/10.5066/P13ISNAI. Producer defined simpl SiMPL Wildlife Magnet Project https://www.sciencebase.gov/catalog/item/6627b9b9d34ea70bd5f01561 Morelli, T. L., Sirén, A. P. K., Patry, R. K., Cliché, R. M., Clark, J. P., Loesberg, J. A., Courtot, K. A., Clarfeld, L. A., Huber, K. E., and Donovan, T. M., 2024, SiMPL Wildlife Magnet Project Data Release Volume 1 (2016 - 2023): U.S. Geological Survey data release, https://doi.org/10.5066/P1VUKJQK. Producer defined usfws Silvio O Conte National Fish and Wildlife Refuge Wildlife Monitoring Project https://www.sciencebase.gov/catalog/item/6634eeced34ee921b712b7e4 Cliche, R.M., Sirén, A.P.K., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, Silvio O Conte National Fish and Wildlife Refuge Wildlife Monitoring Project (2014 - 2024): U.S. Geological Survey data release, https://doi.org/10.5066/P18M87XP. Producer defined vfwd Vermont Fish and Wildlife Department https://www.sciencebase.gov/catalog/item/663ce56cd34e77890839e1c8 Gieder, K.D., Bernier, C.A., Royar, K., Siren, A.P.K., Crumley, K., Smith, T.R., Courtot, K., Wilson, T.L., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, Vermont Fish and Wildlife Department Volume 1 (2014 - 2022): U.S. Geological Survey data release, https://doi.org/10.5066/P14MFBJT. Producer defined wmnf USDA White Mountain National Forest https://www.sciencebase.gov/catalog/item/66520b7dd34e702fe87490d4 Prout, L.S., Siren, A.P.K., Callahan, C.B., Wilson, T.L., Clarfeld, L.A., Huber, K.E., and Donovan, T.M., 2024, USDA White Mountain National Forest Volume 1 (2014 - 2024): U.S. Geological Survey data release, http://doi.org/10.5066/P1PUEYQK. Producer defined nacti North American Camera Trap Images https://lila.science/datasets/nacti Tabak MA, Norouzzadeh MS, Wolfson DW, Sweeney SJ, VerCauteren KC, Snow NP, Halseth JM, Di Salvo PA, Lewis JS, White MD, Teton B. Machine learning to classify animal species in camera trap images: Applications in ecology. Methods in Ecology and Evolution. 2019 Apr;10(4):585-90. Producer defined filename The filename of the image from the original dataset Producer Defined The unique file name of each trail camera image (un-cropped). locationid Name of each location Producer Defined A unique ID for each location in each project x_min Minimum x coordinate (0 to 1 scale) of the animal bounding box used to create image crop. Producer Defined 0.0 0.9907 x_max Maximum x coordinate (0 to 1 scale) of the animal bounding box used to create image crop. Producer Defined 0.00677 1.0 y_min Minimum y coordinate (0 to 1 scale) of the animal bounding box used to create image crop. Producer Defined 0.0 0.9846 y_max Maximum y coordinate (0 to 1 scale) of the animal bounding box used to create image crop. Producer Defined 0.01666 1.0 conf_score Confidence score of the predicted bounding box (0 to 1 scale) Producer Defined 0.101 0.993 taxon Taxon present in the image from the following list of target taxa: The final class list includes American Marten, Bird sp., Black Bear, Bobcat, Coyote, Domestic Cat, domestic cow, Domestic Dog, Fisher, Gray Fox, Gray Squirrel, Human, Moose, Mouse sp., "no-species", Opossum, Raccoon, Red Fox, Red Squirrel, Skunk, Snowshoe Hare, White-tailed Deer, wild boar, and Wild Turkey. ("no-species" indicates the absence of any target species) Producer Defined The taxa above are represented by the following taxonomic serial numbers (itis.gov): American Marten - 180559 Bird sp. - 174371 Black Bear - 180544 Bobcat - 180582 Coyote - 180599 Domestic Cat - 183798 domestic cow - 183838 Domestic Dog - 726821 Fisher - 1086061 Gray Fox - 180609 Gray Squirrel - 180175 Human - 180092 Moose - 898198 Mouse sp. - 180130 "no-species" - (indicates the absence of any target species) Opossum - 179921 Raccoon - 180575 Red Fox - 180604 Red Squirrel - 180166 Skunk - 180562 Snowshoe Hare - 180112 White-tailed Deer - 180699 wild boar - 180722 Wild Turkey - 176136 crop_filename The unique file name, used during training, assigned to the cropped image. Producer Defined The unique file name of each cropped trail camera image. phase Training phase of data sample for model training. Producer Defined 0 training set Producer defined 1 validation set Producer defined 2 testing set Producer defined U.S. Geological Survey - ScienceBase mailing address

Mail Stop 302

Denver CO 80225 1-888-275-8747 sciencebase@usgs.gov Unless otherwise stated, all data, metadata and related materials are considered to satisfy the quality standards relative to the purpose for which the data were collected. Although these data and associated metadata have been reviewed for accuracy and completeness and approved for release by the U.S. Geological Survey (USGS), no warranty expressed or implied is made regarding the display or utility of the data for other purposes, nor on all computer systems, nor shall the act of distribution constitute any such warranty. Digital Data https://doi.org/10.5066/P1E7NDAF None 20250317 Laurence A. Clarfeld Vermont Cooperative Fish and Wildlife Research Unit Post-doctoral Fellow mailing and physical

University of Vermont

Burlington VT 05405 802-656-3011 lclarfel@uvm.edu FGDC Biological Data Profile of the Content Standard for Digital Geospatial Metadata FGDC-STD-001.1-1999