Mining implicit information in rock art via band selection and hyperspectral image unmixing

Rock art records the changes of ancient history and the achievements of human wisdom, making its content highly significant for studying human activities. The extraction of implicit information from rock paintings aims to utilize digital technology to reveal and restore faded rock art, thereby providing additional evidence for the study of ancient civilizations. However, current implicit information extraction methods based on Hyperspectral Imagery (HSI) face challenges such as the loss of critical information, insufficient utilization of spectral and spatial data, and reliance on human expertise. To address the unique characteristics of rock art, this study proposes a implicit information extraction strategy based on band selection and spectral unmixing. The strategy consists of two main steps: band selection and spectral unmixing. The first step employs a target-oriented band selection method to enhance the visibility and separability of implicit information by filtering bands containing the target features. The second step utilizes a spectral blind unmixing network that integrates spatial-spectral characteristics, enabling end-to-end training for pure endmember separation and abundance estimation. Additionally, the strategy incorporates an adaptive spatial information extraction module and a dynamic noise suppression module to effectively handle high-noise scenarios. Experimental results demonstrate that, compared to existing techniques, this method extracts more comprehensive and clearer implicit information from rock art, exhibiting superior performance.