Performance Evaluation of UAV-Coordinated Multi-Scenario Disaster Relief Operations Based on ResNet and Attention Mechanism

Utilizing coordinated UAV formations for emergency disaster relief is a key future trend, but traditional evaluation methods have three major drawbacks: high computational complexity, heavy reliance on expert experience, and poor generalization in multi-scenario small-sample settings. To address these issues, this paper first designs a four-level evaluation index system that covers 5 core capabilities and targets 4 typical disaster relief scenarios. Next, it establishes an AHP model that quantifies the performance of 406 UAV formations, thereby providing high-quality labeled data for subsequent research. Furthermore, the paper constructs a ResNet + Atten deep learning network with a hybrid architecture, which improves both the self-learning ability of expert knowledge and the efficiency of multi-scenario evaluation. To solve small-sample overfitting and expert bias, the paper proposes a physically meaningful controllable perturbation data augmentation method: one that works by perturbing 23 UAV performance metrics within a 5–15% range to expand the sample size. Comparative experiments are conducted using three methods, BP neural networks, ResNet, and LSTM, and results show that ResNet + Atten achieves superior performance. Additionally, the data augmentation method effectively enhances the generalization ability of the model. The proposed method provides a reliable method for evaluating the performance of UAV multi-scenario collaborative disaster relief operations.