First-Order Cross-Domain Meta Learning for Few-Shot Remote Sensing Object Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Remote sensing images present domain complexity due to multi-source sensor variances, challenging conventional cross-domain few-shot methods.
• Existing methods often assume simple distribution shifts, failing to address the heterogeneity inherent in remote sensing data.

Purpose of the Study:

• To propose a novel first-order Cross-Domain Meta Learning (CDML) framework for few-shot remote sensing object classification.
• To address the challenges posed by intrinsic domain complexity and heterogeneity in remote sensing data for few-shot learning.

Main Methods:

• CDML employs a dual-stage domain adaptation task, comprising cross-domain meta-training (CDMTrain) and cross-domain meta-testing (CDMTest).
• CDMTrain utilizes inner-loop multi-domain few-shot task sampling for a teacher model to capture discriminative features and inter-domain distributional divergence.
• A learnable affine transformation model is proposed to adaptively fine-tune update directions, mitigating conflicts in multi-domain scenarios.

Main Results:

• The proposed CDML method demonstrates superior performance compared to state-of-the-art methods across five remote sensing classification benchmarks.
• The alternating cyclic learning paradigm effectively captures genuine domain shifts, guiding the model towards balanced multi-domain performance.
• Iterative domain adaptive task learning in CDMTest progressively enhances domain generalization capability.

Conclusions:

• CDML offers a robust solution for few-shot object classification in remote sensing, adeptly handling complex domain shifts.
• The method's ability to learn from diverse sensor data and adapt to unseen domains signifies a significant advancement in the field.
• The proposed approach provides a strong foundation for future research in cross-domain learning for remote sensing applications.