Abstract:Domain adaptation (DA) strives to mitigate the domain gap between the source domain where a model is trained, and the target domain where the model is deployed. When a deep learning model is deployed on an aerial platform, it may face gradually degrading weather conditions during operation, leading to widening domain gaps between the training data and the encountered evaluation data. We synthesize two such gradually worsening weather conditions on real images from two existing aerial imagery datasets, generating a total of four benchmark datasets. Under the continual, or test-time adaptation setting, we evaluate three DA models on our datasets: a baseline standard DA model and two continual DA models. In such setting, the models can access only one small portion, or one batch of the target data at a time, and adaptation takes place continually, and over only one epoch of the data. The combination of the constraints of continual adaptation, and gradually deteriorating weather conditions provide the practical DA scenario for aerial deployment. Among the evaluated models, we consider both convolutional and transformer architectures for comparison. We discover stability issues during adaptation for existing buffer-fed continual DA methods, and offer gradient normalization as a simple solution to curb training instability.

What problem does this paper attempt to address?

The paper primarily aims to address the issue of progressive domain shift faced by drones or satellites when encountering gradually deteriorating weather conditions during mission execution. Specifically, the paper focuses on the increasing domain gap between the training data and the data encountered during actual operations as weather conditions worsen (e.g., from clear skies to heavy snow). To tackle this problem, the paper proposes the following key points: 1. **Generating Benchmark Datasets**: The paper generates four new benchmark datasets based on two existing aerial image datasets (AID and UCM). These datasets simulate two types of gradually deteriorating weather conditions—cloud cover and snowfall. 2. **Evaluating Continual Domain Adaptation Models**: The paper evaluates the performance of three domain adaptation models, including a baseline standard domain adaptation model (Continual-SHOT) and two continual domain adaptation models (ConDA and UCL-GV). These models can only access a small portion of the target data during testing and need to continually adapt on a single data stream. 3. **Stability Improvements**: The study finds that existing cache-based continual domain adaptation methods have stability issues during the adaptation process. It proposes gradient normalization as a simple yet effective solution to improve the stability of models during continual adaptation. 4. **Exploring Different Network Architectures**: In addition to the traditional ResNet-50 architecture, the paper also explores the application of attention-based Vision Transformer (ViT) and Swin Transformer architectures to evaluate the impact of stronger backbone networks on continual domain adaptation. Through these efforts, the paper aims to provide effective solutions for the domain adaptation problem of aerial images under gradually changing weather conditions in real-world environments.

Continual Domain Adaptation on Aerial Images under Gradually Degrading Weather

Continual domain adaptation on aerial images under gradually degrading weather

Domain Adaptation for Satellite-Borne Multispectral Cloud Detection

ConDA: Continual Adaptation in Remote Sensing Via Visual Style Playback

Continual Test-time Domain Adaptation via Dynamic Sample Selection

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

Delving into the Continuous Domain Adaptation

Self-supervised Domain-agnostic Domain Adaptation for Satellite Images

Complementary Domain Adaptation and Generalization for Unsupervised Continual Domain Shift Learning

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

Test-Time Adaptation for Depth Completion

Domain Adaptation in 3D Object Detection with Gradual Batch Alternation Training

Gradient Regularized Contrastive Learning for Continual Domain Adaptation

Continual Domain Adversarial Adaptation via Double-Head Discriminators

Gradual Domain Adaptation: Theory and Algorithms

DA-RAW: Domain Adaptive Object Detection for Real-World Adverse Weather Conditions

Progressive Domain Adaptation for Change Detection Using Season-Varying Remote Sensing Images

Partial Domain Adaptation for Scene Classification From Remote Sensing Imagery

Attention-based Adversarial Partial Domain Adaptation

Recent Advances in Domain Adaptation for the Classification of Remote Sensing Data

Synthetic Data with Global-Local Level for Adversarial-based Domain Adaptation