Towards Noisy-Reference Paragraph Change Captioning: A Weakly Supervised Detection-Guided Historical-Cascaded Network for Remote-Sensing Images

Remote-sensing image change captioning (RSICC) generally assumes correctly paired bitemporal images and focuses on isolated sentence descriptions, limiting applicability to archives where the correct temporal counterpart may be unknown. We introduce the Noisy-Reference Paragraph-Level Change-Captioning (NRP-CC) task, which requires retrieving the post-change observation from a noisy candidate pool before generating a coherent multi-statement paragraph. We construct the SECOND-ParCap dataset with image-level semantic annotations and paragraph-level change descriptions, and propose a Weakly Supervised Detection-Guided Historical-Cascaded Network (WSDHC-Net). WSDHC-Net leverages Dual Similarity Matching (DS-Matching) for retrieval from noisy candidate pools and employs weakly supervised detection to mine semantic cues for the Historical-Cascaded Paragraph Refinement Change-Captioning (HCPR-CC) module, which generates multi-statement paragraphs describing scene changes. We further introduce SA-ROUGE to evaluate sentence-level correspondence across the generated statements. On SECOND-ParCap, WSDHC-Net achieves Acc@1 scores of 99.53%, 99.29%, and 98.52% under appearance-hard retrieval with 5, 10, and 20 candidates, respectively. In a separate captioning benchmark using identical frozen retrieved inputs across all compared methods, WSDHC-Net attains ROUGE-L and SA-ROUGE scores of 0.5017 and 0.5016, respectively, under reference-assisted sentence-level evaluation; the corresponding paragraph-level scores are 0.2749 and 1.9305. These results support the feasibility of integrating temporal-correspondence retrieval with paragraph-level change description in NRP-CC.