Work

HTC-HAD: A Hybrid Transformer-CNN Approach for Hyperspectral Anomaly Detection

Year: 2025

Type: article

Abstract: Hyperspectral anomaly detection (HAD) identifies anomalies by analyzing differences between anomalies and background pixels without prior information, presenting a significant challenge. Most existing studies leverage the high correlation in spectral and spatial dimensions, primarily focusing on local spectral and spatial information for background reconstruction while neglecting long-range dependencies. This local perception constrains models from fully capturing intrinsic spatial-spectral connections. To address this, we propose a novel hybrid transformer-CNN network for HAD (HTC-HAD). Specifically, HTC-HAD combines CNNs with transformers, where the CNN focuses on local modeling, and the transformer addresses long-range modeling. This dual approach ensures the accurate reconstruction of backgrounds by capturing both local and long-range dependencies. Meanwhile, to reduce model complexity and redundancy among neighboring bands, we incorporate a simplified and effective band selection strategy as preprocessing. Additionally, to prevent anomalies from being reconstructed during background estimation, we employ an adaptive weight loss function to suppress them. Experimental results on several real datasets, both qualitatively and quantitatively, demonstrate that our HTC-HAD achieves satisfying detection performance.

Source: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing

DOI: 10.1109/jstars.2025.3559079

Authors: Minghua Zhao, Wen Zheng, Jing Hu

Institution: Xi'an University of Technology

Language: English

Topic: Remote-Sensing Image Classification

Subfield: Media Technology

Field: Engineering

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,800

Funders: National Natural Science Foundation of China, China Scholarship Council, Natural Science Foundation of Shaanxi Province