Cross-Domain Remote Sensing Image Semantic Segmentation via Iterative Adaptation

The patent divides the target domain adaptation process into multiple iterative stages. In each iteration, the model first performs inter-domain adaptation from source to target domain, then performs intra-domain adaptation within the target domain using generated pseudo-labels. This segmented approach allows the model to progressively improve accuracy without requiring complete model redesign, balancing performance improvement with manageable complexity increments.

The patent generates pseudo-labels for target domain images before formal model training. These preliminary labels are created using the pre-trained source domain model and are used to prepare the target domain data in advance. This preliminary action enables the model to adapt to the target domain more effectively during subsequent training iterations, improving segmentation accuracy while maintaining a systematic training process.

The patent implements a self-service mechanism where the model generates its own training labels through pseudo-labeling. The system uses the model's own predictions on target domain images to create training labels, eliminating the need for manual pixel-level annotation. This self-service approach maintains high segmentation accuracy while dramatically reducing the time and labor costs associated with traditional manual labeling methods.

The patent creates pseudo-labels by copying the labeling process through automated model prediction rather than manual annotation. Instead of humans manually labeling each pixel, the system uses the trained model to generate label copies that can be used for further training. This copying mechanism preserves the quality of labeled data needed for accurate segmentation while avoiding the time-consuming manual labeling process.

The patent goes beyond simple parameter adjustments like gamma correction by implementing iterative domain adaptation that fundamentally changes the model's parameter representation across domains. The system adaptively adjusts model parameters through multiple training iterations, learning domain-specific features and transformations that go far beyond basic brightness or contrast adjustments, thereby achieving high performance despite significant domain differences.

The patent implements a dynamic training process where the model continuously adapts through iterative inter-domain and intra-domain adaptation stages. Rather than using static data augmentation, the system dynamically adjusts to domain shifts by retraining with pseudo-labeled data in multiple iterations, allowing the model to evolve its parameters progressively to handle significant distribution differences while maintaining practical processing efficiency.

The patent segments the domain adaptation process into two distinct phases: inter-domain adaptation (from source to target domain) and intra-domain adaptation (within target domain across different iterations). This segmentation allows the system to address both inter-domain shift and intra-domain variations systematically. By dividing the complex adaptation task into manageable phases, the patent improves segmentation quality while organizing the training process complexity in a structured manner.

The patent implements periodic iterative training where the model cycles through inter-domain adaptation and intra-domain adaptation stages multiple times. In each iteration, pseudo-labels are generated and used to refine the model, creating a periodic improvement cycle. This periodic action allows the model to progressively capture intra-domain differences across multiple cities or regions, improving segmentation quality through repeated refinement while maintaining a systematic training framework.