Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models

TL;DR

This paper introduces Delta-LLaVA, a specialized multimodal large language model for remote sensing change detection, supported by a new comprehensive benchmark, Delta-QA, that advances multi-temporal interpretation capabilities.

Contribution

The work presents Delta-LLaVA with novel modules for change detection and a new benchmark, unifying pixel-level segmentation and question answering in remote sensing.

Findings

Delta-LLaVA outperforms existing models in change deduction accuracy.

The Delta-QA benchmark enables comprehensive evaluation of multi-temporal remote sensing tasks.

Proposed modules effectively isolate and amplify visual differences across time.

Abstract

While Multimodal Large Language Models (MLLMs) excel in general vision-language tasks, their application to remote sensing change understanding is hindered by a fundamental "temporal blindness". Existing architectures lack intrinsic mechanisms for multi-temporal contrastive reasoning and struggle with precise spatial grounding. To address this, we first introduce Delta-QA, a comprehensive benchmark comprising 180k visual question-answering samples. Delta-QA unifies pixel-level segmentation and visual question answering across bi- and tri-temporal scenarios, structuring change interpretation into four progressive cognitive dimensions. Methodologically, we propose Delta-LLaVA, a novel MLLM framework explicitly tailored for multi-temporal remote sensing interpretation. It overcomes the limitations of naive feature concatenation through three core innovations: a Change-Enhanced Attention module that systematically isolates and amplifies visual differences, a Change-SEG module utilizing Change Prior Embedding to extract differentiable difference features as input for the LLM, and Local Causal Attention to prevent cross-temporal contextual leakage. Extensive experiments demonstrate that Delta-LLaVA decisively outperforms leading generalist MLLMs and specialized segmentation models in complex change deduction and high-precision boundary localization, establishing a unified framework for earth observation intelligence.