Deep-Learning-Based Aerial Image Classification for Emergency Response Applications Using Unmanned Aerial Vehicles

Christos Kyrkou, Theocharis TheocharidesWorkshops @ CVPR 2019 · DOI

Abstract

This paper introduces ERNet, a lightweight convolutional neural network (CNN) designed specifically for real-time disaster scene classification on-board Unmanned Aerial Vehicles (UAVs). By optimizing for embedded hardware, the authors achieve performance comparable to heavy-duty models like VGG16 but with a 200x reduction in memory and a 3x speedup on mobile platforms, enabling autonomous emergency response in areas without internet connectivity.

Introduction

In disaster scenarios—such as floods, fires, or building collapses—UAVs are invaluable for rapid situational awareness. However, most current deep learning solutions rely on streaming video to a powerful ground station or the cloud for processing. This creates a critical bottleneck: disaster zones often have destroyed or non-existent communication infrastructure. To be truly effective, a UAV must process images on-board and in real-time. The challenge is that standard high-accuracy CNNs are too computationally expensive and memory-intensive for the low-power embedded processors (like the Odroid-XU4) typically found on small UAVs.

Methodology: The ERNet Architecture

The authors developed ERNet (Emergency Response Network) by systematically testing architectural choices that balance accuracy against hardware constraints. Key design strategies include:

  • Reduced Cost of First Layer: Instead of many small filters, they used a smaller number of larger ($5 \times 5$) filters to capture more spatial information early on with less computation.
  • Early Downsampling: Using max-pooling layers early in the network to aggressively reduce the resolution of feature maps, cutting down the number of operations.
  • Fully Convolutional Architecture: They replaced traditional, memory-heavy "Fully Connected" layers at the end of the network with Global Average Pooling. This drastically reduces the number of parameters and the resulting file size.
  • Residual Connections and Separable Convolutions: Borrowing from ResNet, they added skip connections to help the network learn more effectively despite its compact size. They also utilized depthwise separable convolutions to further reduce mathematical complexity.
The ERNet architecture overview, highlighting the layer configurations, filter sizes, and modular blocks like the Residual and Conv blocks.

The AIDER Dataset

To train and evaluate the model, the authors introduced a new benchmark: the Aerial Image Dataset for Emergency Response (AIDER). It contains four disaster classes (Collapsed Building/Rubble, Fire/Smoke, Flood, Traffic Accident) and one Normal class.

Comprising approximately 2,545 images collected from web sources, news, and UAV flights, the dataset is intentionally imbalanced to reflect real-world conditions. The authors employed heavy data augmentation—including rotations and brightness shifts—to ensure the model's robustness to varying weather and lighting conditions.

Sample images from the AIDER dataset illustrating the five classes: fire/smoke, flood, collapsed buildings, traffic accidents, and normal scenes.

Key Results and Performance

The authors compared ERNet against industry standards like VGG16, ResNet50, and MobileNet on both a PC and an embedded Odroid-XU4 platform.

  • Accuracy: ERNet achieved 90.1% average accuracy, nearly matching the much larger VGG16 (91.9%).
  • Memory Footprint: ERNet requires only 0.3 MB, compared to VGG16’s 59.3 MB and ResNet50’s 96.4 MB. This allows it to fit entirely within the fast cache of many specialized processors.
  • Embedded Performance: On a DJI Matrice 100 UAV equipped with an Odroid-XU4, ERNet achieved ~9 FPS. While lower than PC speeds, it is 3x faster than other state-of-the-art models on the same hardware, making real-time autonomous navigation feasible.

Conclusion

ERNet demonstrates that high-accuracy disaster classification is possible on low-power embedded hardware. By focusing on memory efficiency and architectural optimization, the researchers have enabled UAVs to operate autonomously in communication-denied environments. Future work will focus on incorporating temporal video data and multi-modal sensing, such as infrared data, to further enhance reliability in complex disaster zones.

Summary updated

BibTeX
@inproceedings{kyrkou2019deep,
  author = {Christos Kyrkou and Theocharis Theocharides},
  title = {{Deep-Learning-Based Aerial Image Classification for Emergency Response Applications Using Unmanned Aerial Vehicles}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  doi = {10.1109/cvprw.2019.00077},
}
Figure