Integrating YOLOv8-based disaster detection with deep reinforcement learning for closed-loop UAV emergency response scheduling: A simulation study
- 1School of Traffic and Transportation, Lanzhou Jiaotong University, Lanzhou, China.
Abstract
Aerial perception and dispatch scheduling in unmanned aerial vehicle emergency response are commonly arranged as a one-directional pipeline, and the contribution of closing the loop between them has not been measured under controlled conditions. This study evaluates a framework in which a YOLOv8 scene classifier gates task discovery, a Disaster Situation State Matrix aggregates spatially indexed detection outputs under exponential temporal decay, and a proximal policy optimization scheduler acts on that matrix. Evaluation used a 32 × 32 grid testbed with four vehicles, ten training seeds per configuration, and one hundred frozen evaluation scenarios per seed, with a complete two-by-two ablation separating policy access to the state matrix from the rate at which the matrix is refreshed. Runtime perception updating raised task completion from 0.4971 to 0.5859 relative to an initial-snapshot policy, a gain of 8.87 percentage points (95% CI 7.26 to 10.32, Holm-adjusted p = 0.0078). Policy access to the state matrix carried a main effect of 36.42 percentage points (95% CI 34.83 to 37.79) and refresh rate a main effect of 1.81 percentage points (95% CI 0.90 to 2.82), with no evidence of interaction. Under a shared six-action interface and identical perception constraints, the proposed scheduler exceeded a greedy heuristic by 3.44 percentage points (95% CI 2.03 to 4.91, Holm-adjusted p = 0.0078). Replacing confusion-matrix sampling with replayed YOLOv8 predictions on held-out images reduced completion by 1.28 percentage points, and temporally persistent detection errors extended the longest uninterrupted error run to seventy-one steps. Structural exposure of the state matrix to the scheduling policy, rather than refresh frequency alone, accounts for the observed performance of the framework.
Related
- A sequential integrating virtual simulation and authentic video for preschool teacher candidates’ observational skills training: A design-based intervention study
- GT-PNID: A lightweight graph-temporal network integrating GCN and transformer for power network intrusion detection
- A Multimodal Machine Learning Framework Integrating Genomic, Epidemiological, Environmental, and Clinical Imaging Data for Accelerated Measles Detection and Transmission Inference
- Deep learning for mass extinction detection on fossilized phylogenies: power, limitations and lessons for simulation-based birth–death inference
- Benchmarking circRNA Detection Tools from Long-Read Sequencing Using Data-Driven and Flexible Simulation Framework
BibTeX
@article{zhang2026integrating,
title = {{Integrating YOLOv8-based disaster detection with deep reinforcement learning for closed-loop UAV emergency response scheduling: A simulation study}},
author = {Chunsheng Zhang and Ruichun He},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0361374},
url = {https://doi.org/10.1371/journal.pone.0361374},
note = {CC BY},
}
RIS
TY - JOUR
TI - Integrating YOLOv8-based disaster detection with deep reinforcement learning for closed-loop UAV emergency response scheduling: A simulation study
AU - Chunsheng Zhang
AU - Ruichun He
PY - 2026
DA - 2026/10/29
JO - PLOS One
DO - 10.1371/journal.pone.0361374
UR - https://doi.org/10.1371/journal.pone.0361374
AB - Aerial perception and dispatch scheduling in unmanned aerial vehicle emergency response are commonly arranged as a one-directional pipeline, and the contribution of closing the loop between them has not been measured under controlled conditions. This study evaluates a framework in which a YOLOv8 scene classifier gates task discovery, a Disaster Situation State Matrix aggregates spatially indexed detection outputs under exponential temporal decay, and a proximal policy optimization scheduler acts on that matrix. Evaluation used a 32 × 32 grid testbed with four vehicles, ten training seeds per configuration, and one hundred frozen evaluation scenarios per seed, with a complete two-by-two ablation separating policy access to the state matrix from the rate at which the matrix is refreshed. Runtime perception updating raised task completion from 0.4971 to 0.5859 relative to an initial-snapshot policy, a gain of 8.87 percentage points (95% CI 7.26 to 10.32, Holm-adjusted p = 0.0078). Policy access to the state matrix carried a main effect of 36.42 percentage points (95% CI 34.83 to 37.79) and refresh rate a main effect of 1.81 percentage points (95% CI 0.90 to 2.82), with no evidence of interaction. Under a shared six-action interface and identical perception constraints, the proposed scheduler exceeded a greedy heuristic by 3.44 percentage points (95% CI 2.03 to 4.91, Holm-adjusted p = 0.0078). Replacing confusion-matrix sampling with replayed YOLOv8 predictions on held-out images reduced completion by 1.28 percentage points, and temporally persistent detection errors extended the longest uninterrupted error run to seventy-one steps. Structural exposure of the state matrix to the scheduling policy, rather than refresh frequency alone, accounts for the observed performance of the framework.
SN - 1932-6203
ER -