Greenhouse imagery and deployment telemetry for a containerised edge vision node on a Raspberry Pi 5

Research data supporting the study "Deployment measurements of a containerised edge vision node on a Raspberry Pi 5: thermal, power and latency characteristics, and the attribution of container overhead".

This record contains the greenhouse imagery from which the classifier was trained and the complete per-frame telemetry on which every measurement in the accompanying article rests. Analysis code, container definitions and the edge application that consume these files are at https://github.com/xiaolin200206/edge-container-deployment-measurement

CONTENTS

Background.zip — 640 images, distant outdoor scenes from beyond the greenhouse
Basil_healthy.zip — 498 images, healthy basil foliage collected on site
Real_disease.zip — 560 images, foliar disease symptoms observed on site, in six subdirectories (fungal infection, leaf curl, senescent withering, mealybugs, leaf miner, mite damage)
telemetry.zip — 6.9 MB of deployment logs: four controlled profiling runs, the live field session and an extended continuous-inference session
DEPOSIT_README.md — full documentation of every file, column and caveat; read this first

THE IMAGERY WAS COLLECTED IN ONE PLACE

An enclosed commercial hydroponic basil greenhouse in Cyberjaya, Selangor, Malaysia, over three weeks under natural and highly variable illumination. All captures are hand-held. The deployed node instead uses a fixed camera at a different working distance, so the classifier is trained on one viewpoint and deployed against another.

The three classes are not constructed symmetrically. Background is distant outdoor scenery, distinguished from the other two by scene scale and context rather than by object identity; it is not a generic "everything else" class. Healthy and Disease are both close-range foliage, and Disease spans six visually distinct categories against Healthy's single appearance. This asymmetry matters for the out-of-distribution analysis in the article.

FOUR THINGS TO KNOW BEFORE USING THE TELEMETRY

Rows exist only during active periods. The sleep intervals of the duty cycle appear as gaps in the timestamps, so no idle power baseline can be derived from these logs.

The Bus_ columns measure the Type-C supply input to the whole assembly, not the board's 5 V rail, which the supply module does not instrument. These are node input power figures, inclusive of two cascaded conversion stages, and are not comparable with SoC-level power benchmarks.

The unit of analysis is the run, not the frame. Frames within a run share thermal state, illumination and scene content and are strongly autocorrelated; the effective sample size is two per condition, not the ~100,000 rows each run contains.

The Throttled column is populated by vcgencmd, which the container image does not ship. It reads "No" in the two native runs and "Unknown" in the two containerised ones. "Unknown" means uninstrumented, not clear.

THE PROXY IMAGERY IS NOT DEPOSITED

The study's secondary training condition padded the Disease class with 280 cross-domain images obtained in March 2026 from a publicly available plant-disease collection. The specific source collection was not recorded at the time and could not be identified afterwards, so its licence cannot be established and the images are not redistributed here. This record cannot grant rights that cannot be shown to exist.

The consequence is stated in the article and repeated here: the Real-Only condition is exactly reproducible from this deposit; the Real + Proxy condition is not. A replication does not need these particular images — it needs any cross-domain disease set of matched size, and the released training scripts read whatever is placed in a Proxy_disease/ directory.

REPRODUCIBILITY

Every table and figure in the article and its supplement is regenerated from these logs by the released scripts. analysis/recompute.py emits every quantity stated in the text. The pipeline has been verified to run from a clean checkout against this archive alone and to reproduce those values exactly. The one exception is Figure 10, a screen capture of two delivered alerts whose originating run was not retained; its values are read from the alert payloads and are identified as such in the article.

LICENCE

All files in this record are released under CC BY 4.0. No third-party imagery is redistributed here; every image deposited was captured by the author at the study site.