Cross-layer Attention Enhanced Remote Sensing Image Change Captioning via Mamba-Transformer Interaction

The purpose of Remote Sensing Image Change Captioning (RSICC) is to describe changes between bi-temporal Remote Sensing (RS) images of the same scene using natural language, which requires models to recognize and interpret objects of various sizes and their temporal correlations across different scenarios. Existing methods still face challenges in balancing three aspects: global modeling, computational burden, and effective utilization of channel information. To address these issues, this paper proposes MTI-CC network (Remote Sensing Image Change Captioning via Mamba-Transformer Interaction) for RSICC. MTI-CC integrates both Transformer and Mamba architectures. To effectively utilize spatial and channel dimension information generated at different layers, we employ Cross-layer Change Awareness (CCA) module for iterative processing, generating spatial (channel) attention weights to guide deeper layers in capturing change features. Additionally, we propose a Temporal Aggregation Net (TAN) to capture temporal correlation information through feature complementarity. Comparative experiments with existing encoder-decoder based RSICC methods demonstrate the superior performance of our approach.