Application of Self2Self self-supervised denoising framework to airborne gamma-ray spectrometry data

Airborne gamma-ray spectrometry(AGRS)data are susceptible to degradation due to factors such as flight altitude variations, constraints on detector volume, and intricate background interferences, which collectively hinder the acquisition of high-quality reference samples in field environments and curtail the performance of conventional denoising techniques. This study investigates the potential feasibility of the Self2Self self-supervised denoising algorithm in this domain. The approach requires only a single noisy image, employing random Bernoulli masks to produce incomplete image variants, thereby forcing an enhanced U-Net architecture to reconstruct the values of masked pixels. Furthermore, it integrates dropout mechanisms for multiple sampling alongside ensemble prediction strategies to alleviate variance and generate high-fidelity denoised outputs. Experiments were conducted utilizing data from the LSS experimental zone in Gansu Province, with evaluations performed via peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and feature peak signal-to-noise ratio (FPSNR) metrics. The outcomes reveal that post-denoising, the PSNR values for thorium (Th), potassium (K), and uranium (U) attained 32.03 dB, 30.53 dB, and 28.66 dB, respectively, accompanied by SSIM values of 0.99, 0.97, and 0.91. Additionally, the FPSNR for the characteristic peaks of each radionuclide exhibited enhancements ranging from 2.43 to 2.60 dB, with the uranium series representing a relative improvement of 29.6 %. This methodology proficiently preserves spectral peak morphologies and gradient transitions while effectively attenuating stripe artifacts, thereby augmenting overall image quality. The research furnishes novel perspectives for leveraging AGRS in mineral prospecting and environmental surveillance, substantiating the robustness of the Self2Self framework in processing datasets characterized by low count rates and stochastic noise perturbations.