Link-adaptive edge-cloud inference for UAV-based plastic mulch residue assessment

Abstract

Plastic mulch residue assessment from UAV imagery is jointly constrained by segmentation accuracy, end-to-end latency and uplink bandwidth, since onboard inference loses boundary fidelity on fragmented film while cloud inference is throttled by the link, and these constraints have been treated separately in prior work. This study develops a link-adaptive edge-cloud inference framework combining a three-factor decision module that fuses EWMA-smoothed bandwidth, entropy-derived confidence and motion-compensated residual into a single score assigning each frame to local termination, region-of-interest offloading or full-frame offloading; a plastic-mulch-aware student-teacher architecture transferring representations across a two-thousand-fold scale gap from Sentinel-2 imagery to sub-centimetre UAV imagery; and NSGA-II optimisation recovering the latency-accuracy Pareto front rather than a scalar compromise. Evaluated on a self-collected UAV residue dataset under emulated 4G-LTE, 5G-NR and Wi-Fi 6 links, the balanced operating point attains an IoU of 0.826, recovering 97.1% of the cloud-only SegFormer-B3 accuracy at a median end-to-end latency of 72 ms, and responds to abrupt link transitions 3.4 times faster than fixed-threshold offloading. Accuracy degrades on fragments below 200 pixels, corresponding to about 50 cm 2 at the acquisition scale, which bounds the operational value of the framework to residue within the range that mechanised collection engages.