Deep Reinforcement Learning for Adaptive Soil Monitoring in AIoT‐Enabled Precision Agriculture
Concurrency and Computation: Practice and Experience · 2026-11 · 38 (22)
DOI: 10.1002/cpe.70987 ↗Kurz gefasst
Fachartikel in der Zeitschrift Concurrency and Computation: Practice and Experience, online erschienen am 08.11.2026. Verfasst von Enes Bajrami und 2 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben.
Abstract (Originalsprache)
Artificial Intelligence of Things (AIoT) monitoring systems must balance sensing quality, communication delay, information freshness, and energy consumption under changing operating conditions. This study presents an action‐dependent Deep Reinforcement Learning framework for adaptive agricultural sensing and communication management. A chronological dataset of 2000 soil, atmospheric, vegetation, and baseline 5G‐latency observations supplies the exogenous environmental sequence, while explicitly defined simulation equations generate the endogenous queue, packet‐delivery, battery, energy, realized‐latency, and Age‐of‐Information states. At each decision step, the policy jointly selects continuous sensing intensity, transmission power, and payload compression, thereby affecting both immediate performance and subsequent system states. Proximal Policy Optimization (PPO), Soft Actor‐Critic (SAC), and Twin Delayed Deep Deterministic Policy Gradient (TD3) are compared with an explicitly defined deterministic rule‐based policy across five training seeds, three workload levels, and training checkpoints of 25,000, 50,000, and 100,000 steps. At the final checkpoint, SAC achieved the highest monitoring fidelity and packet‐delivery ratio under all workloads, including 77.3% fidelity and 92.6% delivery under high workload. PPO achieved the lowest high‐workload energy consumption at 0.84 units per step, while SAC achieved the lowest high‐workload P95 latency at 10.03 ms. The rule‐based policy remained competitive under low workload but accumulated 41.2 queued packets, reached 1.54‐step mean information age, and obtained an episode return of 10.0 under high workload. Reward ablation further confirmed that monitoring fidelity and action‐dependent latency terms materially influence the learned policies. The results demonstrate that continuous DRL control improves robustness under demanding workloads, although no algorithm dominates every monitoring and resource objective.
Zitieren
APA
Bajrami, E., Idrizi, H., & Idrizi, F. (2026). Deep Reinforcement Learning for Adaptive Soil Monitoring in AIoT‐Enabled Precision Agriculture. Concurrency and Computation: Practice and Experience, 38(22). https://doi.org/10.1002/cpe.70987
Vancouver
Bajrami E, Idrizi H, Idrizi F. Deep Reinforcement Learning for Adaptive Soil Monitoring in AIoT‐Enabled Precision Agriculture. Concurrency and Computation: Practice and Experience. 2026;38(22). doi:10.1002/cpe.70987
BibTeX
@article{Bajrami2026,
author = {Bajrami, Enes and Idrizi, Hirijete and Idrizi, Florim},
title = {Deep Reinforcement Learning for Adaptive Soil Monitoring in AIoT‐Enabled Precision Agriculture},
journal = {Concurrency and Computation: Practice and Experience},
year = {2026},
volume = {38},
number = {22},
doi = {10.1002/cpe.70987}
}
EUVIMED