RADLERMachidon, Alina Luminita (Avtor) Krašovec, Andraž (Avtor) Igreţ, Ioana C. (Avtor) Pejović, Veljko (Avtor) Machidon, Octavian-Mihai (Avtor) adaptive inferenceenergy-efficient deep learningimage segmentationprecision agricultureslimmable neural networksunmanned aerial vehiclesweed detectionReal-time UAV weed detection must operate under tight onboard compute and energy constraints, yet most lightweight segmentation pipelines still commit the system to a single fixed accuracy-efficiency operating point. This article introduces RADLER, a context-aware model selection strategy for slimmable neural networks that predicts, for each input image, the smallest network width that remains close to the best attainable segmentation quality. RADLER combines image-level contextual features, feature scaling with redundancy diagnostics, and threshold-controlled optimal-width labels to expose multiple operating points rather than a single adaptive configuration. We evaluate the method on two public UAV weed detection datasets, compare it against static-width baselines and an oracle reference, and characterize runtime and energy behavior on an NVIDIA Jetson Nano. Across the evaluated settings, RADLER matches full-width performance within uncertainty in one setting and otherwise trades a small, measurable IoU decrease for lower average selected width, enabling energy savings between 25% and 50% depending on the operating point. The results show that context-aware slimmable model selection can make onboard agricultural vision pipelines more transparent in their trade-offs and better suited to battery-constrained UAV platforms.20262026-08-26 13:42:15Članek v reviji186076UDK: 004.93:004.8:632.51ISSN pri članku: 2169-3536DOI: 10.1109/ACCESS.2026.3718374COBISS_ID: 288258307sl