Change-LISA: Language-Guided Reasoning for Remote Sensing Change Detection

Remote sensing change detection (RSCD) is widely used for monitoring land-surface dynamics. Most models, however, output an unconditional “all-changes” map and do not account for user intents expressed in natural language. This can lead to spurious masks under counterfactual (i.e., negative/ mismatched) queries. This paper studies language-guided reasoning change detection. Given a bitemporal image pair and a text instruction, the model outputs an instruction-conditioned change mask; the correct output can be an empty mask when the queried change is absent. To support this setting, ReasonRS-110K is built as a benchmark of 112,317 image–instruction–mask triplets. Instructions are organized into hierarchical reasoning levels (L1–L3) together with L0 negative queries for compliance evaluation. Change-LISA is proposed as a parameter-efficient framework that combines a frozen Siamese SAM encoder and an explicit difference/fusion module with a multimodal LLM (LLaVA). The LLM emits atoken whose embedding prompts a lightweight mask decoder. Only LoRA adapters and the decoder are optimized. On ReasonRS-110K, Change-LISA reaches 63.1 IoU and 77.3 F1 on positive-query levels (L1–L3) and 95.4 NCA with 0.12 FPA on L0 negative queries, outperforming representative CD-only, RIS, and MLLM baselines. The code for this work is available at https://github.com/jxy0413/changeLisa.