Answers are grounded in this paper and cite their sources.
Vision-based disease detection is increasingly deployed on low-cost edge hardware in agriculture, yet accuracy reported on curated datasets can diverge sharply from field behaviour, and the thermal behaviour of the host device is rarely characterised at all. This paper reports the deployment of a durian (Durio zibethinus) disease-detection system on a Raspberry Pi 5 and examines three factors that determine its field reliability. First, a dataset-integrity audit combining exact (MD5) and perceptual-hash duplicate detection across the train/validation boundary with per-class instance-to-image tracing identifies a train/validation leak in the early-stage dataset: one byte-identical pair and 117 near-duplicate pairs, resolving to 22 source components under a provenance-based grouping, and traced to an augmentation-before-split workflow. Re-evaluation on a validation partition with the implicated images removed shows the leak inflated the affected class by 35.0% in AP@0.5 and aggregate mAP@0.5 by 9.5%. Second, eleven training runs across three architecture families — including a transformer-based, anchor-free detector with three times the parameter count — leave two classes at or near zero average precision, while bootstrap resampling of the validation set, applied to the two original-submission checkpoints, shows 93% of the nominal difference between them to originate in a single validation instance. Retraining on a source-level leak-free partition recovers one of those classes from 0.000 to 0.399, which is consistent with its zero having been a property of the validation sample rather than of the class, and motivates a candidate three-way distinction between feature-space, label-space and evaluation-space failure. All detection figures are computed on a 126-image validation partition holding one and two instances in two classes; they are not estimates of general detection performance, and no independent source-level test partition was available. Third, a duty-cycle thermal characterisation across five configurations, replicated four weeks apart, shows that the operative failure mode on this platform is not throughput loss but protective shutdown: continuous inference triggered 99 and 108 software thermal cut-offs across the two rounds and a 15 s sleep interval 41 and 3, whereas a 30 s interval recorded none in either round under the tested conditions, with 45 s retaining the larger thermal margin. Throttling itself cost only 2.6% of median inference latency (419.0 to 408.0 ms), so what duty-cycle scheduling protects is thermal margin and uninterrupted operation rather than frame rate. Because each cut-off suspends inference for a fixed interval, continuous operation delivers 95.5% rather than the 100% coverage its schedule nominally promises: for an operator walking an inspection round, 8.2 min of the 183 min run examine no tree, with no indication of when. Direct power measurement at the system-on-chip domain shows the exchange is not free in the other direction either: duty cycling lowers mean SoC power by 28.2% but raises SoC energy per inference by 39.4%, because the idle floor is paid throughout the sleep interval. The audit protocol, the failure taxonomy and the cut-off-based thermal criterion are stated as procedures that could be applied to other resource-constrained agricultural deployments; their generality is not tested here, since the evidence is one durian dataset, one Raspberry Pi 5 and one software stack.
No takes yet. Share an insight, caveat, or question.
Lin Ding Shan (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: