Greenhouse imagery of basil (Ocimum basilicum) for on-device disease classification: site-collected and cross-domain proxy sets
Year: 2026
Type: other
Abstract: Research data supporting the study "Deployment measurements of a containerised edge visionnode on a Raspberry Pi 5: thermal, power and latency characteristics, and the attributionof container overhead", submitted to Computers and Electrical Engineering. This record holds both the image dataset from which the deployed classifier was trained andthe complete deployment telemetry on which every measurement reported in that study rests.The analysis code, container definitions and edge application that consume these files areat https://github.com/xiaolin200206/unified-agtech-engine. IMAGE DATA The imagery was collected over three weeks in a commercial greenhouse cultivating basil(Ocimum basilicum), under natural and highly variable illumination. Background.zip 640 images. Non-plant and mixed scene content of the kind encountered by a fixed camera observing the canopy. Basil_healthy.zip 498 images. Healthy basil foliage, site-collected. Real_disease.zip 560 images. Foliar disease symptoms observed on site. Proxy_disease.zip 280 images. Cross-domain disease imagery sourced from open-access agricultural image databases, kept as a separate archive because it was not collected at the study site. Provenance and the licence of the source collection are given in the README. The classifier distinguishes three classes: Background, Healthy and Disease. Two trainingconditions are compared in the accompanying study. Under Real-Only the Disease class isdrawn entirely from Real_disease. Under Real+Proxy it is sample-size-matched at 560 images,half from Real_disease and half from Proxy_disease, so that data composition varies whiledataset size does not. The matched design isolates the effect of composition from theconfounding effect of size. The three site-collected archives are original camera captures made in situ. The proxy setoriginates from a different acquisition regime and differs from the site imagery inresolution and capture conditions as well as in symptom morphology; that difference is thedomain shift the accompanying study evaluates. All images are resized to 224 x 224 duringtraining, which affects the two sources asymmetrically: the site captures are substantiallydownsampled while the proxy images are close to their native resolution. TELEMETRY telemetry.zip contains the deployment logs in three groups. profiling_runs/ Four controlled runs of three hours each under a 60 s active / 15 s sleep duty cycle: two executing natively on the host and two inside a Docker container, interleaved across day and night so that ambient temperature is balanced between conditions. Each run directory holds per-frame telemetry (inference latency, processor and memory utilisation, SoC temperature, throttling status, predicted class and confidence, and supply voltage, power and battery state sampled once per second over I2C), per-cycle event markers at every active-to-sleep transition, and a record of the run conditions. field_log/ The live greenhouse session: 5,989 frames over 181 s of continuous in-situ operation. supplementary_session/ An extended continuous-inference session: 313,011 frames over 3 h 33 min under a 180 s / 45 s duty cycle. Telemetry rows exist only during active periods; the sleep intervals appear as gaps. The supply figures are node input power measured at the Type-C connector, upstream of the conversion to the board's 5 V rail, which the supply module does not instrument. NOTE ON THE PROFILING COMPARISON The two profiling conditions differed in interpreter, inference-runtime and base-imageversion as well as in execution mode: the native runs used Python 3.13 with ONNX Runtime1.29.0 on Debian 13, the containerised runs Python 3.9 with ONNX Runtime 1.19.2 on Debian11, eighteen releases and roughly 23 months apart. The measured differences are propertiesof the two configurations as built and are not attributable to containerisation alone. Thisis treated at length in the accompanying study and in the repository's CHANGELOG. LICENSING The site-collected image archives (Background.zip, Basil_healthy.zip, Real_disease.zip) andtelemetry.zip are released under CC BY 4.0. Proxy_disease.zip is redistributed under theterms of its original source collection, stated in the README; it is not covered by theCC BY 4.0 licence applied to the rest of this record.
Author: Lin, Ding Shan
Institution: UCSI University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7213757517