On-Device Deployment of Optuna-Optimized YOLOv8n via Flutter for Real-Time and Explainable Cocoa Pod Disease Detection

DOI

10.2991/978-94-6239-805-4_35How to use a DOI?

Keywords

Cocoa Disease Detection; Edge AI; Mobile Deployment; TensorFlow Lite; YOLOv8

Abstract

Deep learning-based cocoa pod disease detection models achieve high accuracy on GPU-accelerated infrastructure, yet are rarely validated for on-device deployment in the bandwidth-constrained smallholder plantations where they are most needed. This study deploys a previously published Optuna Bayesian-TPE-optimized YOLOv8n checkpoint, selected for zero-false-negative Macro Recall and lower inference latency over a marginally higher-mAP Hyperband alternative, into a Clean-Architecture Flutter application via TensorFlow Lite, providing the first paired, on-device validation of a model previously evaluated only in its pre-conversion form. On an identical 30-image held-out test set, TFLite conversion and mobile-pipeline integration reduce mAP@0.5 from 0.984 to 0.908 (-7.6 percentage points), with degradation concentrated in the Black Pod Rot class (-18.4 pp) and correlated with close-range, large-bounding-box framing in the affected images. On-device inference on a mid-range Android device averages 4,991.5 ms/frame (≈0.20 FPS, ≈657× slower than training-time GPU latency), falling well below conventional real-time thresholds, a gap we quantify as the dominant deployment cost rather than obscure. A fully randomized, replicated 8-condition stability protocol found no evidence of memory leaks (one-way ANOVA, F(2,34)=0.104, p=0.90) and no OS-reported thermal throttling across sessions up to 90 minutes. The application further integrates a fully offline, farmer-facing disease-guidance module surfacing etiology, prevention, and treatment information directly from detection results, adapted from a published phytopathological reference, though its in-application adaptation remains pending independent domain-expert re-validation; this integration is demonstrated architecturally but not empirically validated with end-users. These results characterize concrete, previously unmeasured costs of translating an optimized detection model into field-deployable mobile software.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - I Putu Oka Wisnawa
AU  - I Made Dwi Jendra Sulastra
AU  - Putu Manik Prihatini
AU  - I Made Agus Oka Gunawan
AU  - Ni Luh Putu Listya Dewi
AU  - Ni Wayan Widyasari Damayanti
PY  - 2026
DA  - 2026/10/08
TI  - On-Device Deployment of Optuna-Optimized YOLOv8n via Flutter for Real-Time and Explainable Cocoa Pod Disease Detection
BT  - Proceedings of the International Conference on Sustainable Green Tourism Applied Science - Engineering Applied Science 2026 (ICOSTAS-EAS 2026)
PB  - Atlantis Press
SP  - 334
EP  - 346
SN  - 2352-5401
UR  - https://doi.org/10.2991/978-94-6239-805-4_35
DO  - 10.2991/978-94-6239-805-4_35
ID  - Wisnawa2026
ER  -