Abstract:Domain shift caused by, e.g., different geographical regions or acquisition conditions is a common issue in machine learning for global scale satellite image processing. A promising method to address this problem is domain adaptation, where the training and the testing datasets are split into two or multiple domains according to their distributions, and an adaptation method is applied to improve the generalizability of the model on the testing dataset. However, defining the domain to which each satellite image belongs is not trivial, especially under large-scale multi-temporal and multi-sensory scenarios, where a single image mosaic could be generated from multiple data sources. In this paper, we propose an self-supervised domain-agnostic domain adaptation (SS(DA)2) method to perform domain adaptation without such a domain definition. To achieve this, we first design a contrastive generative adversarial loss to train a generative network to perform image-to-image translation between any two satellite image patches. Then, we improve the generalizability of the downstream models by augmenting the training data with different testing spectral characteristics. The experimental results on public benchmarks verify the effectiveness of SS(DA)2.

What problem does this paper attempt to address?

The paper primarily addresses a common issue in satellite image processing known as domain shift and proposes a novel solution. Specifically, due to factors such as different geographical locations, acquisition conditions (e.g., sensor type, time, etc.), satellite images often exhibit spectral variations when processed using machine learning, leading to performance degradation of trained models when faced with new scenarios. To tackle this problem, the paper introduces a method called Self-supervised Domain-agnostic Domain Adaptation (SS(DA)²). ### Main Contributions: 1. **Problem Definition**: The paper first clarifies the concept of Domain Adaptation (DA) and points out that a fundamental premise of traditional DA methods is the need for explicit domain knowledge, i.e., how to divide data into multiple domains based on data distribution. However, in global-scale applications, such domain definitions become very challenging. 2. **Proposed Method**: To address the above issue, the authors propose a DA method that does not require predefined domain divisions—SS(DA)². This method trains a generative network by designing a contrastive generative adversarial loss function to achieve image-to-image translation between any two satellite image patches. Additionally, it enhances the training data to include spectral characteristics of different test sets to improve the generalization ability of downstream models. 3. **Experimental Validation**: The paper conducts experiments on public benchmark datasets, and the results demonstrate the effectiveness of SS(DA)², particularly in outperforming existing image-to-image translation-based DA methods without the need for domain definitions. ### Technical Details: - **Self-supervised Learning**: Utilizes contrastive learning in self-supervised learning techniques to measure the similarity between sample pairs, optimizing the network to make different views from the same image more similar. - **Generative Adversarial Networks**: Combines contrastive learning techniques with Generative Adversarial Networks (GAN) to achieve style transfer between any two image patches without explicitly modeling domain features. - **Contrastive Generative Adversarial Loss**: This is the core loss function of the entire framework, ensuring that the generated images have style characteristics similar to the target images. In summary, this research aims to develop a general and downstream application-independent DA method capable of adapting to satellite image domains without predefined domain division rules.

Self-supervised Domain-agnostic Domain Adaptation for Satellite Images

Towards Self-Similarity Consistency and Feature Discrimination for Unsupervised Domain Adaptation.

Pseudo Features Guided Self-training for Domain Adaptive Semantic Segmentation of Satellite Images

Unsupervised Domain Adaptation of High-Resolution Aerial Images Via Correlation Alignment and Self Training.

Domain Adaptation for Remote Sensing Image Semantic Segmentation: An Integrated Approach of Contrastive Learning and Adversarial Learning

Active Domain Adaptive Semantic Segmentation with Regional Relative Entropy for Remote Sensing Images

SiamSeg: Self-Training with Contrastive Learning for Unsupervised Domain Adaptation in Remote Sensing

S&GDA: An Unsupervised Domain Adaptive Semantic Segmentation Framework Considering Both Imaging Scene and Geometric Domain Shifts

Self-training and Multi-level Adversarial Network for Domain Adaptive Remote Sensing Image Segmentation

A domain adaptation network for cross-imaging satellites sar image ship classification

Unsupervised Domain Adaptation for Remote Sensing Image Semantic Segmentation Using Region and Category Adaptive Domain Discriminator

Self-Training-Based Unsupervised Domain Adaptation for Object Detection in Remote Sensing Imagery

Self-training guided disentangled adaptation for cross-domain remote sensing image semantic segmentation

MultiDAN: Unsupervised, Multistage, Multisource and Multitarget Domain Adaptation for Semantic Segmentation of Remote Sensing Images

Two-Stage Domain Adaptation Based on Image and Feature Levels for Cloud Detection in Cross-Spatiotemporal Domain

Decomposition-Based Unsupervised Domain Adaptation for Remote Sensing Image Semantic Segmentation

Unsupervised Domain Adaptation With Joint Supervised Sparse Coding And Discriminative Regularization Term

Semi-Supervised Domain Adaptation via Adaptive and Progressive Feature Alignment

DAST: Unsupervised Domain Adaptation in Semantic Segmentation Based on Discriminator Attention and Self-Training.

Semi-supervised Domain Adaptation via Joint Transductive and Inductive Subspace Learning