Source-linked AI summary
EmergencyNet: Efficient Aerial Image Classification for Drone-Based Emergency Monitoring Using Atrous Convolutional Feature Fusion
Christos Kyrkou, Theocharis Theocharides
TL;DR
UAV emergency monitoring needs aerial image classification that remains accurate under onboard latency, memory, and power constraints. The paper introduces a dedicated dataset, compares existing approaches, and proposes EmergencyNet, whose lightweight atrous-convolution design delivers near-state-of-the-art accuracy with substantially higher performance.
Problem
Onboard UAV classification must operate despite limited computation, low power, and potentially unavailable cloud communication in disaster areas.
Method
The paper builds a dedicated emergency-response aerial image database, compares CNN approaches, and proposes EmergencyNet using depthwise atrous convolutions for multiresolution features.
Results
EmergencyNet provides approximately 95.7% accuracy and up to approximately 20× faster performance with reduced memory and computation compared with state-of-the-art-oriented models.
Takeaways & Limitations
The resulting accuracy–speed trade-off supports efficient emergency-response monitoring on low-cost, low-power embedded UAV platforms.
Takeaways & Limitations
The system still makes errors when incidents occupy little image area or viewpoints are unsuitable, while further architectural optimization remains possible.
Abstract
from arXiv · showhide
Deep learning-based algorithms can provide state-of-the-art accuracy for remote sensing technologies such as unmanned aerial vehicles (UAVs)/drones, potentially enhancing their remote sensing capabilities for many emergency response and disaster management applications. In particular, UAVs equipped with camera sensors can operating in remote and difficult to access disaster-stricken areas, analyze the image and alert in the presence of various calamities such as collapsed buildings, flood, or fire in order to faster mitigate their effects on the environment and on human population. However, the integration of deep learning introduces heavy computational requirements, preventing the deployment of such deep neural networks in many scenarios that impose low-latency constraints on inference, in order to make mission-critical decisions in real time. To this end, this article focuses on the efficient aerial image classification from on-board a UAV for emergency response/monitoring applications. Specifically, a dedicated Aerial Image Database for Emergency Response applications is introduced and a comparative analysis of existing approaches is performed. Through this analysis a lightweight convolutional neural network architecture is proposed, referred to as EmergencyNet, based on atrous convolutions to process multiresolution features and capable of running efficiently on low-power embedded platforms achieving upto 20x higher performance compared to existing models with minimal memory requirements with less than 1% accuracy drop compared to state-of-the-art models.
1 Introduction
UAV emergency monitoring requires accurate aerial classification that can run autonomously under tight onboard computing and power constraints. The paper introduces EmergencyNet and a dedicated dataset to balance classification accuracy with real-time performance.
- Motivation: UAVs may need onboard processing because disaster areas can lack reliable cloud communication and high-end infrastructure.Limited payload, computational resources, and power make algorithmic efficiency important for real-time hazard detection.
- Problem Definition: The task is to assign semantic labels to UAV-captured aerial images and alert operators to recognized hazards during continuous monitoring.The considered scenario uses a UAV following a predetermined path while analyzing camera frames.
- Proposed Approach: EmergencyNet uses depthwise atrous convolutions to gather multiresolution features while balancing accuracy and inference speed on embedded hardware.The model is designed for deployment on the UAV or its mobile control station.
- Dataset: The study introduces a dedicated emergency-response aerial image database containing more images than existing datasets.The dataset supports aerial image classification for emergency response applications.
- Results: EmergencyNet achieves approximately 95.7% accuracy and up to approximately 20× faster performance while targeting low-cost, low-power devices.The claimed accuracy is near state-of-the-art with reduced computational demands.
- Evaluation: The paper evaluates CNN architectures using accuracy, inference speed, memory, and an experimental UAV setup with embedded and mobile-ground-station processing options.The comparison covers both model characteristics and deployment configurations.
2 Background and Related Work
Prior aerial emergency-classification methods largely rely on deep CNNs, transfer learning, and desktop-class processing. The paper motivates specialized embedded models by reviewing CNN architectures and application-specific disaster-classification approaches.
- CNN Background: CNN architectures learn hierarchical visual representations for image detection, recognition, and segmentation through multilayer training.The background discusses CNNs as the foundation for many computer-vision classification systems.
- CNN Architectures: VGG16, ResNet, Inception, Xception, EfficientNet, MobileNet, SqueezeNet, and ShuffleNet represent different accuracy, scaling, separability, and parameter-efficiency strategies.These architectures vary in depth, filter composition, residual learning, separable convolutions, automated search, and channel processing.
- Emergency-Response Classification: Earlier emergency-classification methods include pixel thresholds, Gaussian mixture models, support vector machines, pretrained CNNs, segmentation, and region-proposal processing.The reviewed applications cover fire, avalanche debris, wildfire, and natural-disaster aerial imagery.
- Prior Results: 91% accuracy was reported for a custom five-class natural-disaster test set, with average processing time below 3 seconds.This result is attributed to a VGG-based transfer-learning approach.
- Deployment Gap: Existing approaches often process UAV footage remotely on desktop-class GPUs, making communication latency and embedded-platform limitations important deployment concerns.The paper positions embedded suitability as a gap in the existing literature.
3 Deep Learning for Aerial Disaster-Event Classification
This section constructs the AIDER dataset and develops EmergencyNet, a lightweight CNN using multiresolution atrous feature fusion for efficient aerial disaster-event classification on embedded platforms.
- 3.1 Dataset Collection: AIDER was constructed because no widely used public dataset existed for emergency-response aerial image classification.It contains manually collected images spanning Fire/Smoke, Flood, Collapsed Building/Rubble, Traffic Accidents, and Normal classes.
- 3.1 Dataset Collection: The dataset groups visually similar disaster imagery into broad classes, while finer-level classification is left for future work.Active flames and smoke, for example, are grouped together.
- 3.1 Dataset Collection: Images were collected from online sources, manually inspected for relevant centered events, standardized in size, and augmented with geometric and photometric transformations.Augmentations include rotations, translations, mirroring, cropping, zooming, illumination changes, and color shifting.
- 3.2 Network Analysis: Pretrained-network transfer learning is compared with custom architectures because large pretrained models can exceed UAV resource constraints and are costly to modify or retrain.The transfer-learning baselines include VGG16, ResNet50, FireNet, MobileNets, Xception, ShuffleNet, EfficientNet, and SqueezeNet.
- 3.3 Atrous Convolutional Feature Fusion (ACFF): The ACFF design fuses depthwise atrous-convolution features across dilation rates, combining representations with maximum, addition, concatenation, or averaging operations.Atrous convolutions enlarge receptive fields without increasing parameter count, while subsequent 1 × 1 convolutions combine channel features.
- 3.4 Macro-Architecture Design Choices: EmergencyNet reduces embedded inference cost through channel reduction, early downsampling, a small first layer, and a pyramid-shaped architecture.The first layer uses 16 filters of size 3 × 3 with stride 2; training also balances classes through majority undersampling and minority oversampling.
4 Experimental Evaluation and Results
The evaluation validates the proposed methodology against existing networks on the developed dataset and tests it in two UAV processing settings.
- The study compares the proposed approach with existing networks on the developed aerial-image dataset.
- Real experiments evaluate embedded processing, where computations run onboard a resource-constrained UAV platform.
- A second setting evaluates remote-based processing, in which the UAV transmits the processing workload remotely.
4.1 Performance Metrics
The evaluation uses frame rate and mean F1 score to assess real-time processing and class-balanced classification performance under nonuniform class distributions.
- Frames per second measures the rate at which each model processes incoming camera frames and is inversely related to per-frame processing time.
- Mean F1 score captures the balance between precision and recall across classes.
- Mean F1 score is used because it handles label imbalance and provides a more appropriate performance measure across classes.
4.2 Overall Performance Analysis and Comparison
EmergencyNet is compared with custom fusion networks and pretrained architectures across accuracy, computational complexity, memory, and classification difficulty. It achieves competitive accuracy with substantially lower resource demands, while errors remain concentrated in ambiguous or poorly framed incidents.
- The standard convolutional network has the highest parameter count and memory demand, whereas depthwise-separable convolution reduces resources but causes a considerable accuracy drop.Spatially separable convolution improves accuracy relative to depthwise-separable convolution but requires additional parameters.
- Add and maximum fusion perform best among ACFF-based networks because they emphasize specific spatial locations without increasing memory or processing requirements.Average fusion may suppress information and reduce important features.
- 95.7% F1 score accuracy is achieved with approximately 360KB memory, about 30× smaller than MobileNetV3.EmergencyNet has comparable accuracy to EfficientNet and MobileNetV3 while using fewer parameters.
- Images are often misclassified when incidents are unclear, occupy little image area, or resemble another class such as flood imagery resembling a road.The confusion matrix identifies normal-class errors as a major improvement opportunity.
4.3 Qualitative Evaluation of Learning
The qualitative evaluation examines which image regions and learned features influence EmergencyNet's classification decisions. It uses averaged and upscaled activations from successive layers to localize high-activation regions.
- Feature-map analysis averages activations at each layer and scales them to the preceding layer's map size.This procedure is used to identify image parts producing the highest activations.
4.4 Embedded Application Results
EmergencyNet is evaluated on embedded and mobile platforms designed for resource-constrained UAV applications. It delivers higher frame rates than competing models, although platform-specific FLOPs do not directly determine CPU performance.
- UAV On-board Processing: Approximately 25 FPS is achieved on the quad-core ARM Cortex-A57 platform, compared with at most approximately 9 FPS for other state-of-the-art models.The platform is described as a compact, energy-efficient processing option suitable for UAVs.
- Qualitative Evaluation: Feature-map visualizations show high activations concentrated on event-specific cues, including fire's red-orange glow and rubble from collapsed buildings.These regions are used by the network in forming its predictions.
- UAV On-board Processing: 2–20× speedup is observed against other models on the embedded evaluation.Further gains may be possible through quantization and bit-reduction techniques.
- Mobile Control Station: Approximately 19 FPS is reached on the tablet mobile control station, while state-of-the-art networks remain below 10 FPS.The tablet processes the UAV camera feed without adding weight or power demands to the UAV platform.
4.5 Discussion and Further Improvements
EmergencyNet combines efficient computation with extensions for higher-resolution imagery and smoother video predictions. The discussion also identifies feature-fusion improvements, difficult data cases, and possible multimodal integrations.
- EmergencyNet matches state-of-the-art accuracy with a fraction of the memory and computation.
- Tiling, batching, and parallelization could preserve comparable frame rates when processing higher-resolution images.
- Video prediction flickering was reduced by weighting predictions using similarity between successive frames.
- Automated model search is identified as a next step for improving the feature-fusion architecture.
- Traffic incidents span diverse scenarios, while limited data collection motivates subclass refinement and synthetic-data generation as research directions.
- Combining EmergencyNet with people, vehicle, infrared, or thermal sensing and geospatial tagging could enhance situational awareness.
5 Concluding Remarks
The paper presents EmergencyNet as an efficient building block for real-time disaster-event recognition on UAVs. Its evaluation reports faster inference, lower memory use, comparable or better accuracy, and a dedicated dataset for future work.
- EmergencyNet provides a trade-off among accuracy, inference speed, and complexity for on-board UAV processing.
- 20× faster inference, an order of magnitude less memory, and similar or better accuracy were reported against existing models.
- The dedicated aerial image dataset is intended to support further advances in emergency-response models and techniques.