Optimized loss and self attention for enhanced domain adaptation in remote sensing image classification

Area of Science:

• Computer Science
• Geospatial Science
• Artificial Intelligence

Background:

• Remote sensing (RS) image classification traditionally requires extensive labeled data, which is costly and labor-intensive to acquire.
• Existing methods struggle with large-scale, high-dimensional data and domain shifts caused by variations in acquisition conditions or sensors.
• Domain adaptation techniques aim to transfer knowledge from labeled to unlabeled domains, but existing approaches have limitations.

Purpose of the Study:

• To develop a unified domain adaptation framework for remote sensing image classification.
• To integrate primary, secondary, and entropy losses with a self-attention mechanism.
• To evaluate the proposed methodology's effectiveness across various state-of-the-art neural network models and datasets.

Main Methods:

• Proposed a novel framework incorporating primary (center, triplet) and secondary (MMD, CORAL, entropy) losses.
• Integrated a self-attention mechanism within the unified framework.
• Evaluated performance on neural network models (VGG, ResNet, AlexNet, GoogLeNet, EfficientNet, MobileNet, ViT) using RSSCN7, NWPU-RESISC45, AID, and UCMerced datasets.

Main Results:

• The integrated framework demonstrated effectiveness in handling domain shifts in remote sensing image classification.
• Systematic review confirmed the performance of various losses on state-of-the-art neural networks.
• Experiments validated the proposed methodology using features from classification and penultimate layers.

Conclusions:

• The proposed attention-integrated domain adaptation framework offers a more robust and accurate solution for remote sensing image classification.
• This approach effectively addresses the challenges posed by distribution variability and the need for unlabeled data.
• The findings pave the way for improved remote sensing systems capable of adapting to diverse and changing environments.