ABSTRACT:

Validation evidence for the GRIME AI feature extraction module. Synthetic test images with known analytical properties are processed through GRIME AI and results are compared against expected values. Includes analytical validation of intensity, Shannon entropy, GCC, and ExG metrics using pure-color images, and directional validation of GLCM, Gabor, LBP, and Fourier descriptors using synthetic image pairs. Contains the validation script, formal test plan, synthetic test images, and CSV results from a validation run conducted on 2026-03-27.

ABSTRACT:

Validation evidence for the GRIME AI image triage module. Synthetic test images with known blur and brightness properties are processed through GRIME AI and results are compared against expected classifications. Includes validation of blur threshold detection, brightness lower and upper bound detection, and combined condition handling across six test cases. Contains the validation script, formal test plan, synthetic test images organized by test case, and CSV results from a validation run conducted on 2026-03-27

ABSTRACT:

This collection holds sample datasets and work products from groups participating in the PEP2026 Workshop. Pixels to Enviro Patterns 2026 (PEP2026), is a two-day workshop at the University of Nebraska–Lincoln (Nebraska), where cameras, AI, and storytelling converge for scientific discovery. Held March 7–8 at Hardin Hall on East Campus, PEP 2026 invites researchers, educators, and creatives to explore open imagery and dive into hands-on learning with GRIME AI software. Collaborate on mini-projects to turn pixels into environmental patterns that drive scientific inquiry and narratives. Whether you’re passionate about water, phenology, artificial intelligence, or communication, this is your chance to connect, create, and be inspired.

Datasets are curated for use with GRIME AI software, which is available as a Conda package (https://anaconda.org/channels/GRIMELab/packages/grime-ai/overview), with source code and Wiki on GitHub (https://github.com/JohnStranzl/GRIME-AI/wiki). GRIME AI is free, open-source software (Apache 2.0). GRIME AI leverages Meta's Segment Anything 2 (SAM2) model for image segmentation and also facilitates the entire data science workflow, from data retrieval to model deployment.

This work was completed utilizing the Holland Computing Center of the University of Nebraska, which receives support from the UNL Office of Research and Innovation, and the Nebraska Research Initiative.

ABSTRACT:

Validation evidence for green chromatic coordinate (GCC) calculations produced by GRIME AI (GaugeCam Remote Image Manager Educational Artificial Intelligence). Synthetic test images with known color properties were generated using a Python script producing two image types used in the GCC comparison: solid-color images with uniform fill (flat) and solid-color images with a stipple effect applied via Gaussian-blurred random noise (stippled). Both types span 65 green hues across an 8 x 8 grid of HSV saturation and value levels plus one pure green [0, 255, 0] image, at a resolution of 1024 x 768 pixels saved as JPEG. The script additionally produces splotch images — circular green patches composited onto a clay-colored background in flat and stippled permutations — which are included in the deposit but were not used in the GCC comparison analyses. GCC values were independently computed by GRIME AI and by the PhenoCam Network pipeline (Northern Arizona University) and manually compiled into a spreadsheet for comparison.

GCC values computed by GRIME AI were compared against GCC values computed by the PhenoCam Network pipeline, used here as a benchmark reference, across both flat and stippled solid image sets. Agreement between the two implementations was near-perfect: flat images yielded a mean absolute difference of 0.0% (SD = 0.00003) and stippled images yielded a mean absolute difference of 0.0% (SD = 0.0002). These statistically insignificant discrepancies are attributed to minor floating-point rounding differences introduced by CPU architecture variation during image processing rather than any methodological inconsistency, confirming that GRIME AI produces reproducible and accurate greenness calculations consistent with an established community standard.

A secondary analysis examined the effect of image serialization on GCC consistency by comparing GRIME AI calculations performed on images held in memory against calculations performed on those same images after saving to disk and reloading. Results were evaluated separately for flat images (mean difference = 0.02%, SD = 0.00383) and stippled images (mean difference = 0.03%, SD = 0.00344). Both differences are traced to JPEG lossy compression, which introduces small per-channel rounding errors at the pixel level (e.g., a pure green pixel [0, 255, 0] becomes [0, 255, 1] after a save-reload cycle). The slightly larger variability in stippled images is consistent with greater spatial heterogeneity providing more opportunity for compression artifacts to accumulate. The dataset includes the image generation script, all synthetic test images, and a spreadsheet recording GCC outputs from both GRIME AI and PhenoCam alongside computed differences.