Comparative YOLO Analysis for UAV Disaster Surveillance
Generational comparison of lightweight YOLO nano-scale classification models and activation function micro-optimizations for real-time UAV edge-cloud disaster response on the AIDER dataset.
Hardware & Software Stack
Research Publications & Citations
Problem & Solution
The Challenge
Deploying deep learning models on Unmanned Aerial Vehicles (UAVs) is highly critical for immediate situational awareness in disaster zones (such as locating flooded areas, fires, or structural collapses). However, UAVs operate under severe edge computing constraints: limited battery life and onboard hardware resources require models with extremely low latency and small parameter footprints.
While new model architectures like YOLOv11 and YOLOv12 are released regularly, their adoption introduces complex computational layers that may not translate to real-world accuracy gains under aerial constraints, environmental noise, and varying lighting conditions.
The Approach
I designed a quantitative comparative benchmark of three generations of lightweight nano models: YOLOv8n-cls, YOLOv11n-cls, and YOLOv12n-cls on the Aerial Image Database for Emergency Response (AIDER) dataset.
Furthermore, I investigated micro-architectural optimization by swapping model activation functions dynamically between ReLU and SiLU (Sigmoid Linear Unit).
The benchmarking revealed a critical counter-intuitive finding: YOLOv8n-SiLU achieved the highest peak accuracy of 98.37% (with validation loss of 0.0587), outperforming the newer YOLOv12n (97.82%) and YOLOv11n-SiLU (97.67%), showing that precise activation tuning on stable, older baselines is superior to simply adopting newer generation backbones.
Interactive UAV Inference Simulator
Select an aerial disaster scene, configure the model backbone and activation optimization, adjust flight variables (altitude, camera tilt, lighting), and run simulated inference to test confidence drops, latency variations, and live terminal logging as it would execute on an onboard UAV computing module.
Onboard Logic Analysis
Select an AIDER class and click "Run Onboard Inference" to evaluate.
Experimental Methodology & Code
The research utilized a dataset of 6,433 aerial images from the AIDER database. We split the dataset into 70% training (4,501 images), 20% validation (1,286 images), and 10% testing (646 images). All images were preprocessed and resized to 640x640 pixels to fit the YOLO input graph.
To evaluate activation micro-optimization, I wrote a recursive modules traversal script in PyTorch. The script descends down the model's computational graph and replaces every occurrence of nn.SiLU with nn.ReLU, creating the modified variants:
import torch.nn as nn
def replace_activation(module):
for name, child in module.named_children():
if isinstance(child, nn.SiLU):
setattr(module, name, nn.ReLU()) # Swap SiLU to ReLU
else:
replace_activation(child) # Recursively traverse submodules
# Apply the replacement directly to the loaded network
replace_activation(model.model)
print("Model activations micro-optimized to ReLU.")SiLU (Sigmoid Linear Unit)
SiLU applies a smooth, non-monotonic gating activation: \(f(x) = x \cdot \sigma(x)\) where \(\sigma(x) = \frac{1}{1 + e^{-x}}\). The smooth gradient improves convergence flow but has exponential overhead on edge CPUs.
ReLU (Rectified Linear Unit)
ReLU applies a simple linear threshold: \(f(x) = \max(0, x)\). It is computationally cheap (zero floating-point operations) but suffers from dead neurons and poor gradient propagation on complex high-resolution aerial datasets.
Comparative Performance Benchmarking
All models were trained under identical conditions: 50 epochs, Stochastic Gradient Descent (SGD) optimizer, dynamic Automatic Mixed Precision (AMP) enabled, on an NVIDIA T4 GPU. The checkpoint weights yielding the lowest validation loss were selected for final test set evaluation.
Table I: Model Performance Matrix on AIDER
| Model Generation | Activation Function | Top-1 Accuracy | Validation Loss | Best Epoch |
|---|---|---|---|---|
| YOLOv8n (Baseline) | ReLU (Modified) | 97.59% | 0.1063 | 41 |
| YOLOv8n-SiLU (Best) | SiLU (Default) | 98.37% | 0.0587 | 47 |
| YOLOv11n | ReLU (Modified) | 97.51% | 0.0942 | 41 |
| YOLOv11n-SiLU | SiLU (Default) | 97.67% | 0.0988 | 42 |
| YOLOv12n | ReLU (Modified) | 97.82% | 0.0853 | 37 |
| YOLOv12n-SiLU | SiLU (Default) | 97.05% | 0.0999 | 40 |
Critical Research Insights
Ablation: Generational SiLU Impact
Applying the smooth, non-monotonic SiLU activation yielded significant accuracy gains on older backbones: +0.78% for YOLOv8n and +0.16% for YOLOv11n. However, forcing SiLU on YOLOv12n led to a performance drop of -0.77% (down to 97.05%). This suggests that YOLOv12's native attention mechanisms and re-parameterized blocks are highly self-calibrated; adding extra activation gates over-complicates the network.
Confusion Matrix Analysis
Across all architectures, class sensitivities remained above 95% except for the "Traffic Incident" class, which was consistently capped at 94%, with 6% of cases misclassified as "Normal". This error stems from visual similarities in low-altitude drone perspectives: high-density car flow on highways and road layouts look highly identical to a static car accident scene, requiring higher resolution mapping or temporal video feeds to resolve.