Unsupervised Domain Adaptation Semantic Segmentation of High-Resolution Remote Sensing Imagery With Invariant Domain-Level Prototype Memory

Semantic segmentation is a key technique involved in automatic interpretation of high-resolution remote sensing (HRS) imagery and has drawn much attention the community. Deep convolutional neural networks (DCNNs) have been successfully applied to HRS semantic task due their hierarchical representation ability. However, heavy dependency on large number training data with dense annotation sensitiveness variation distribution severely restrict potential application DCNNs for imagery. This study proposes novel unsupervised domain adaptation network (MemoryAdaptNet) MemoryAdaptNet constructs an output space adversarial learning scheme bridge discrepancy between source target narrow influence shift. Specifically, we embed invariant feature memory module store domain-level context information because features obtained from only tend represent variant current limited inputs. integrated by category attention-driven aggregation pseudo further augmenting pixel representations. An entropy-based label filtering strategy used update high-confident images. Extensive experiments under three cross-domain tasks indicate that our proposed remarkably superior state-of-the-art methods.