Change detection is a critical task in remote sensing Earth observation for identifying changes in the Earth’s surface in multi-temporal image pairs. However, due to the time-consuming nature of image collection, labor-intensive pixel-level labeling with the rare occurrence of building changes, and the limitation of the observation location, it is difficult to build a large, class-balanced, and diverse building change detection dataset, which can result in insufficient changed sample pairs for training change detection models, thus degrading their performance. Thus, in this article, given that data scarcity and the class-imbalance issue lead to the insufficient training of building change detection models, a novel multi-temporal sample pair generation method, namely, Image-level Sample Pair Generation (ISPG), is proposed to improve the change detection performance through dataset expansion, which can generate more valid multi-temporal sample pairs to overcome the limitation of the small amount of change information and class-imbalance issue in existing datasets. To achieve this, a Label Translation GAN (LT-GAN) was designed to generate complete remote sensing images with diverse building changes and background pseudo-changes without any of the complex blending steps used in previous works. To obtain more detailed features in image pair generation for building change detection, especially the surrounding context of the buildings, we designed multi-scale adversarial loss (MAL) and feature matching loss (FML) to supervise and improve the quality of the generated bitemporal remote sensing image pairs. On the other hand, we also consider that the distribution of generated buildings should follow the pattern of human-built structures. The proposed approach was evaluated on two building change detection datasets (LEVIR-CD and WHU-CD), and the results proved that the proposed method can achieve state-of-the-art (SOTA) performance, even if using plain models for change detection. In addition, the proposed approach to change detection image pair generation is a plug-and-play solution that can be used to improve the performance of any change detection model.

Referring Remote Sensing Image Segmentation (RRSIS) is a representative multimodal understanding task for remote sensing, which segments designated targets from remote images according to free-form natural language descriptions. However, complex remote sensing characteristics, such as cluttered backgrounds, large-scale variations, small scattered targets and repetitive textures, lead to unstable visual grounding and further spatial grounding drift, resulting in inaccurate segmentation results. Existing approaches typically perform implicit visual–linguistic fusion across encoding and decoding stages, entangling spatial grounding with mask refinement. This tightly coupled formulation lacks explicit structural constraints and is prone to cross-modal ambiguity, especially in complex remote sensing layouts. To address these limitations, we propose a Structurally consistent and Grounding-aware Stagewise Reasoning Framework (SGSRF) that follows a grounding-first, segmentation-second paradigm. The framework decomposes inference into three cascaded stages with progressively imposed structural constraints. First, Cross-modal Consistency Refinement (CCR) lays the foundation for stable spatial grounding by enhancing visual–textual structural alignment via CLIP-based features and Structural Consistency Regularization (SCR), producing well-aligned multimodal representations and reliable grounding cues. Second, Grounding-aware Prompt (GPG) Generation bridges grounding and segmentation by converting aligned representations into complementary sparse and dense prompts, which serve as explicit grounding guidance for the segmentation model. Third, Grounding Modulated Segmentation (GMS) leverages the Segment Anything Model (SAM) to generate fine-grained mask prediction under the joint guidance of prompts and grounding cues, improving spatial grounding stability and robustness to background interference and scale variation. Extensive experiments on three remote sensing benchmarks, namely RefSegRS, RRSIS-D, and RISBench, demonstrate that SGSRF achieves state-of-the-art performance. The proposed stagewise paradigm integrates structural alignment, explicit grounding, and prompt-driven segmentation into a unified framework, providing a practical and robust solution for RRSIS in real-world Earth observation applications.