Improving semantic segmentation accuracy in thin cloud interference scenarios by mixing simulated cloud-covered samples

Abstract

Thin cloud interference presents a significant challenge for the semantic segmentation of optical satellite imagery, which directly degrades the model accuracy and causes difficulties in sample selection. This paper generated a dataset named Populus euphratica and Tamarix chinensis discrimination (PTD), containing both cloudless and thin cloud scenarios. Based on this PTD dataset, an enhanced Atmospheric Scattering Model with Nonlinear Optimization (ASM_NL) was proposed to simulate high-fidelity thin clouds by incorporating two vital nonlinear terms: the point spread function and the Perlin noise. Additionally, we adopt a strategy of mixing simulated thin cloud-covered images (STCI) into the training set at a certain proportion to improve the semantic segmentation accuracy in thin cloud-covered scenarios. The conclusions are as follows: 1) ASM_NL can simulate high-fidelity clouds at an average Jensen-Shannon distance of 0.0699. 2) When dealing with medium- and high-cloud density datasets, mixing STCI proved to be more effective than cloud removal in mitigating thin cloud interference, resulting in average macro F1 score improvements of 0.164 and 0.094, respectively. 3) The semantic segmentation accuracy improved significantly by mixing STCI with a minimal proportion of 1/60, demonstrating the activation of model transfer capabilities. This study provides a concise and efficient methodology for effectively mitigating thin cloud interference in deep learning-based optical satellite imagery analysis.