Change Detection in Earth Observation: From Pixel Differences to Language Descriptions

An expanding fleet of orbital satellites and fine-resolution low-altitude platforms now generates hundreds of petabytes of Earth observation data daily. Processing these vast, multi-dimensional datasets demands resilient algorithms that minimize computational cost while maximizing throughput for change detection. Applications span agricultural and deforestation monitoring, urban sprawl and earthworks, wildfire trajectories, seasonal inundation and flash flood dynamics, shifts in open water bodies, glacier retreat, land subsidence, illegal mining expansion, structural collapse and infrastructure disruption in emergency response, displaced population camps, algae blooms and oil spills, and dark vessel and fishing monitoring, among others.Vision-language models (VLMs) have emerged as a promising alternative, offering cross-modal feature alignment and operational flexibility. They have drawn strong interest from researchers and practitioners and represent a fast-growing area of remote sensing research. This Research Topic reports the latest advances in VLMs applied to change detection and welcomes both theoretical and applied contributions. We expect the resulting body of work to become a reference point for the ongoing advancement of remote sensing image processing and interpretation.Change detection has long been a challenging problem in remote sensing, and recent years have brought a surge of innovative approaches. Traditional pixel-based methods gave way to time-series analysis and trajectory modeling, then to deep learning, computer vision, transformer architectures, and foundation models. Current research emphasizes architectures that capture long-range spatial dependencies and global contextual interactions across multitemporal images, alongside semantic change detection, weakly-supervised and unsupervised learning, explainability, robustness, and multisensor and multimodal fusion across diverse data formats.Critical information now extends beyond pixels: inputs for change detection have evolved from standalone imagery to include textual information. VLMs have emerged as a cutting-edge solution for detecting and describing changes in remotely sensed imagery, well suited to the demands of smart, near real-time applications.Articles may address, but are not limited to, the following topics:• Change captioning and automated narrative generation• Conversational and text-prompted change querying• Multi-modal vision-language foundation models for remote sensing• Zero-shot and open-vocabulary change detection• Reasoned attribution and cause-and-effect analysis• LLM-guided agentic pipelines for geospatial workflows• Domain-specific applications and impact• Cross-modal fusion and SAR-optical dynamic reasoning• Quality control, hallucination reduction, and uncertainty mitigation• Conversational spatiotemporal scenario simulation and forecasting Topic editors: Cláudia Maria Almeida, LAURENT POLIDORI, Shizhen Chang Source: https://www.frontiersin.org/research-topics/85960/change-detection-in-earth-observation-from-pixel-differences-to-language-descriptions