This dataset contains 12,496 (and counting) Images from motion triggered camera traps deployed at NEON domain 20 at Pu'u Maka'ala Natural Area Reserve. This dataset is prelimary, as images will continue to be recieved through 2025. Images contain one or more focal indivudals of NEON monitored understory plants, that can match therefor to NEON data product DP1.10055.001 Plant Phenology Observations. The data also includes various captures of native and non-native birds, as well as some nocturnal mammals.

We deployed 19 low-cost WOSPORTS G100 mini trail cameras across five NEON-monitored species (two individuals per species; 10 plants total) to capture both plant phenology and faunal activity at two vertical strata. These low cost cameras trigger based on passive motion infrared sensor, and are primarily marketed as hunting aids for imaging animals, offering no timed capture function. Ten units were mounted on manually built tripods constructed from metal rods—adjusted to hold the lens at mid-canopy height and angled to frame the focal plant in the foreground—while nine were placed atop standard cinder-block stands at ground level. Each camera was labeled with a unique ID, configured for synchronized capture intervals, and secured in economical, weatherproof housing to withstand outdoor conditions. Each camera used two SD cards, which NEON scientists alternately swapped every 15 days during their routine transect visits. The tripod-mounted cameras recorded seasonal changes in flowers and foliage as well as arboreal visitors, whereas the cinder-block units monitored terrestrial fauna moving around the plant base. Currently 12,484 images were collected from January 27th to April 3, covering a roughly 2 month window at the site.

Supported Tasks and Leaderboards

Currently a subset of the data has been annotated for evaluation of animal detection, across 4 cameras at 2 differnt taxa for a total of 4,625 images. Using the initial combo of megadetector and bioclip to detect animals and filter non bird detections, we acheived a precion of .375 and recall of .76. Further work will evaluate more species detectors and classifiers as well as filtering methods.

Dataset Structure

The dataset is designed for viewing and slicing with the tool [voxel51] (https://voxel51.com/). Fiftyone provides diverse and flexible methods for sorting, slicing, and visualizing datasets and has a backend database. To accomplish many of the same features of premium colaborative versions of 51 we developed a pipeline for dataset creation and app launch from a versioned csv.

For each "dataset" the code to lauch 51 reads from "datasets.json" a set of paths to the csv, the images, and accompanying detection and classification dataframes. Annotations made using the "tags" feature in 51 are automatically saved to the csv at app closure and can be later manually loaded into formalized dataset fields.

datasets.json = {
  "PUUM_pheno_transect_cams": {
    "name": "PUUM_pheno_transect_cams",
    "df": "./data/PUUM_pheno_transect_cams_fo_df_new_data_updated_20250427_154753.csv",
    "dir": "/fs/ess/PAS2136/Hawaii-2025/in-use/Camera-traps/ID*/*/*.jpg",
    "detections": "./data/agg_detections_new.csv",
    "classifications": "./data/agg_classifications_new.csv",
    "fields" : ["camera_id","direction","common_name","mount","year", 
    "month", "day", "hour", "minute", "second",
     "valid_date", "timestamp_check","date_bin"]
  }
}

Within the dataset dir, images from SD cards at field sites are uploaded according to the following structure

Camera-traps
    <ID01-C>/
        <R32_01_A_25_02_25_25_03_12>/
            IMAG0000.jpg
            IMAG0001.jpg
            ...
        <R32_01_A_25_03_26_25_04_03>/
        ...
        <R{size of card}_{camera_id}_{SD card ID (A or B)}_{start date %YY%MM%DD}_{end date %YY%MM%DD}>

    <ID02-T>/
        <R32_02_A_25_02_25_25_03_12>/
    ...

Data is updating using the protocol described in "update_guide.md" in the accompanying project repository.

Data Instances

Images are ~6000 x 4000 pixels, very noisy camera trap images. They contain a black banner along the bottom portion that contains a camera ID code and a timestamp. OCR is used to automatically extract these timestamps from the same region of each image. Images from tripods, folders ending in (T) are set to capture every 5 seconds 1 image, which cameras on cinder blocks capture 3 images per each trigger and have a 15 second cool down period before another trigger occurs.

{dataset_name}fo_df{update_timestamp}.csv : central csv file for all image captures that contain both metadata for camera id as well as processed timestamps and image specific annotations.

{dataset_name}_agg_detections{update_timestamp}.csv : derived csv file with megadetector detections for each sample in the available image folder.

{dataset_name}_agg_classifications{update_timestamp}.csv : derived csv file with top1 predictions for crops produced from running inference following protocol described in accompanying repo.

annotated_detections.csv : derived csv file containing verified positive detections for susbet of dataset, and bioclip features for extracted crops.

deployment_details.csv: google sheet in project folder that contains camera metadata including geolocation and PUUM pheno ID number. Kept seperate to prevent geolocation leakage.

Data Fields

PUUM_pheno_transect_cams_fo_df_{update_timestamp}.csv :

filepath: Full path to the file in the filesystem.
tags: tag strings saved to samples in fiftyone dataset, will be a list of strings
created_at: Date and time when the file was originally created.
last_modified_at: Date and time when the file was last modified. Auto generated by 51, may be deprecated
timestamp: Extracted timestamp from image using OCR, may have errors.
basename: Filename without the directory path.
date_bin: upload period of data, or which folder it corresponds to.
mount: Tripod or cinder block (T/C).
common_name: common name of focal individual plant.
direction: Orientation of camera placement.
camera_id: Unique identifier for the camera.
year: Year extracted from the timestamp.
month: Month extracted from the timestamp.
day: Day extracted from the timestamp.
hour: Hour extracted from the timestamp.
minute: Minute extracted from the timestamp.
second: Second extracted from the timestamp.
date: Full date (year-month-day)associated with the file.
SD_card_GB: Storage capacity (in GB) of the SD card used.
SD_card_id: Identifier for the specific SD card used.
valid_date: Boolean or flag indicating whether the date metadata is considered valid. Mostly if numeric or possible date.
timestamp_check: Boolean flag for checking date falls within expected window and pertains to study period.
folder: Name of the folder containing the file.

PUUM_pheno_transect_cams_agg_detections.csv:

img_id: path to original image on OSC.
bbox: predicted bounding bo by megadetector.
category: 0 animal, 1 human, 2 vehicle.
confidence: model prediction score.

PUUM_pheno_transect_cams_agg_classifications.csv:

file_name: path on OSC to CROP image prodcued from megadetector script.
kingdom: bioclip predicted label.
phylum: bioclip predicted label.
class: bioclip predicted label.
score: bioclip predicted confidence.

annotated_detections.csv:

filepath: Full path to the CROP file on OSC.
date: Full date (year-month-day) extracted from the file.
bbox: Bounding box coordinates for detection in image.
confidence: Confidence score of detection from megadetector.
class: Bioclip pedicted class label for animal in bbox.
og_filepath: Original filepath to full image file.
filepath_y: Alternative or second filepath (often after merging datasets). ...
all metadata fields from fo_df above ...
species: Identified species from manual annotation.
visit_number: bird visit grouped temporaly by a threshold of 15 seconds per camera.
correct: Whether a classification or prediction was correct (True/False).
og_filename: Original filename before any modification.
md_success: confusion matrix category of bioclip detection(FN, TP, FP).
classification_success: confusion matrix category of bioclip classification(FN, TP, FP).
combined_success: confusion matrix category (FN, TP, FP).
feature_0 to feature_511: Individual components of the extracted feature vector from BioCLIP.
unique_id: Unique identifier for the sample.

Deployment_Metadata.csv:

camera level metadata feilds from fo_df.csv +
NEON Species Code: String code for species epithet.
NEON Individual ID Num: NUmber to match to target individaul in NEON's records.

Data Splits

All data has been kept raw, with inference being tested but no seperate splits have been generated.

Dataset Creation

Curation Rationale

This work intended to demonstrate the ability to measure phenological changes and track plant use by animals at a NEON site. See more details in accompanying paper.

Data Collection and Processing

All steps for preprocessing of data are contained in the update_guide.md in the accompanying repo. Preprocessing remains currently light, mainly consisting of OCR of timestamps and flagging of anomolous timestamps or metadata categories.

Who are the source data producers?

This data is captured at the National Ecological Observatiory Network Domain 20 Pu'u Maka'ala Reserve, part of the Hawaii natural area reserves. Cameras were deployed as part of the Imageomics and ABC Center Class on AI and Ecology.

Annotations

Annotation process

Annotations were done at the initail image level (not crop level) using custom protocol in the fiftyone app. Annotatiosn are recorded as string tags that can be applied to sample images and will be saved in the fo_df.csv. 4 cameras were chosen to be filtered for missed detections and false positives. ID02 and ID26 contain tripod images of two Ohelo or native blueberry that flowered during the study period, capturing many visitors. ID09 and ID12 are centralized on a berry producing pukiawe, that experienced oma'o and kalij pheasant visits. All images across the intial studu period were reveiwed for these cameras, with detection tags of "no_detection_0.2" (indicating Megadetector threshold) and "false_positive_detection" corresponding to false negative and false positive bird detections. When the image had also been classified by BioCLIP to be bird triggered, false positive tags extended to the classification success as well. For false negatives and true positives, species labels were also labeled using the tags feature. Overall, the annotated detections csv contains ~300 verified positive detections and classificaitons, as well as 54 manually added false negative detections.

Who are the annotators?

Luke Meyers

Personal and Sensitive Information

This data does contain endangered plants, and threatened birds, as well as numerous images of field technicians. Before release all identities should be obscured and geographic data should be confirmed to be unextractable from the image exifdata.

Considerations for Using the Data

This data does have some issues that are the result of it being preliminary, as well as an educational experience.

PUUM_pheno_transect_cams_fo_df_{update_timestamp}.csv :

many timestamps need to be corrected. Ideally a script should prompt a human for help and format input into standard format. Currently ~200 images have erroneous timestamps.
in the folders on OSC, many duplicate or erroneous uploads of data from NEON are present. These are artifacts from the rsync process and are not included in metadata dataframes, but should be noted.

annotated_detections.csv:

with the inclusion of false negatives, the originally crop oriented dataframe shifted to include samples without a crop_path, classification prediction or features. Og_filepath is the new main sample identifier column linking to the images from the camera traps HOWEVER,
confusion resulted from multiple detections in the same image, some crops became artificial incorrect predictions to assignment of labels from other crops in the same frame. These samples were rectified but processess may need to be updated in the future. At the time of cropping, a new 51 dataset should be created and false negatives can be created by hand and noted as such? (open for comment)

Bias, Risks, and Limitations

Data exhibits classic long tailed distribution. Data is noisy and captures moving plants at irregular intervals. Cameras are mounted alongside transect, an access road which we expect to be a deterrint to shyer birds and will most likely strongly impact our distributions of species. Geolocation data on deployment details metadata contains species at risk of being poached, and should be NOT be released.

Recommendations

Consider collection bias when attempting bioligical analysis or comparison with other methods.

Licensing Information

Image data is currently private.

Citation

BibTeX:

Acknowledgements

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Glossary

More Information

Dataset Card Authors

Luke Meyers, Anirudh Potlapally

Dataset Card Contact

Feel free to shoot either of us an email... I'm at [email protected]

Downloads last month: 5

Curated by:

Luke Meyers, Anirudh Potlapally

Repository:

]

Paper: