Memory-Based Pruning of Deep Neural Networks for IoT Devices Applied to Flood Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Monitoring

Background:

• Automatic flood detection is crucial for mitigating socioeconomic impacts, with river images offering a readily available data source.
• Current deep neural network (DNN) models for flood detection are computationally intensive, limiting their deployment on resource-constrained Internet of Things (IoT) devices.

Purpose of the Study:

• To develop a DNN architecture pruning algorithm for creating memory-efficient models for flood detection.
• To enable the deployment of flood detection systems on low-capability IoT devices with unstable internet connectivity.

Main Methods:

• Proposed a novel deep neural network (DNN) architecture pruning algorithm.
• Algorithm identifies a pruned DNN version within a user-defined memory footprint.
• Evaluated the pruned models on a memory-constrained wireless sensor network (WSN) for urban river flood detection.

Main Results:

• The proposed algorithm successfully generated pruned DNN models matching specified memory constraints.
• Pruned models exhibited minimal to no degradation in segmentation performance compared to original models.
• Deployed pruned models in a WSN for urban river flood detection demonstrated competitive results.

Conclusions:

• The DNN pruning algorithm effectively reduces model size while preserving performance for flood detection tasks.
• This approach facilitates the deployment of advanced computer vision models on edge devices for real-time environmental monitoring.
• The developed method offers a viable solution for robust and efficient flood detection in resource-limited WSNs.