Distributed Deep Learning With Gradient Compression for Big Remote Sensing Image Interpretation.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral images (HSIs) are crucial for Earth observation but present interpretation challenges due to high dimensionality.
• Deep neural networks (DNNs) are effective for target detection in HSIs, but massive data volumes strain edge device capabilities.
• Existing methods struggle with the computational demands of hyperspectral target detection (HTD) on Internet of Things (IoT)/edge devices.

Purpose of the Study:

• To introduce a decentralized deep learning approach for efficient HTD on edge devices.
• To address the communication bottleneck in distributed learning for HSI analysis.
• To develop a gradient compression technique that maintains accuracy while reducing overhead.

Main Methods:

• Proposed distributed background learning, a decentralized deep learning strategy for HTD.
• Introduced gradient compression via centroid (GCC) to compress gradients and reduce communication.
• Tested the framework on large hyperspectral datasets using a Ring All-reduce distributed system.

Main Results:

• Distributed background learning demonstrated superior speed for HTD compared to single-node approaches.
• GCC achieved 50% gradient compression with minimal accuracy loss (0.01%) for target detection.
• The method significantly reduced communication overhead, outperforming existing gradient compression techniques.

Conclusions:

• Distributed background learning with GCC is a feasible and efficient solution for HTD on edge devices.
• The framework effectively balances computational requirements, communication efficiency, and detection accuracy.
• This approach is expected to accelerate the adoption of distributed training for IoT/edge-based remote sensing.