Exploring Infrared and Visible Feature Fusion for Earth Surface Anomaly Detection: From Benchmark Dataset to Spatial-Spectrum Adaptation Network

Earth surface anomaly detection (ESAD) is a valuable and challenging research topic in the field of remote sensing. Earth surface anomalies (ESAs) occur infrequently, evolve rapidly, and demonstrate varied spectral characteristics, which poses huge challenges to existing ESAD methods. In this paper, we propose a novel ESAD network based on the vision-language architecture, called ESA-CLIP, to strengthen spectral feature representation of various ESAs. Specifically, an enhanced spatial-spectrum adapter (ES2Adapter) is developed to capture robust visual feature while promoting cross-band feature interaction. Meanwhile, a dynamic gate fusion (DGFusion) is further constructed to fuse visible (RGB) and infrared (IR) spectral features effectively, thereby learning spectrum-sufficient feature representation. Moreover, we build a multispectral Earth surface anomaly detection dataset (MS_ESAD) to address the scarcity of existing ESA samples. This dataset collects 70,000 Sentinel-2 multispectral images and involves six spectral bands and nine anomaly categories, providing foundational data support for ESAD research. Extensive ablation and comparative experiments on this dataset demonstrate the competitive performance of our ESA-CLIP compared to state-of-the-art scene classification methods and its potential towards general ESAD. The code and dataset will be available at: https://github.com/WangLibo1995/ESA_CLIP.