Evaluating Sentinel-2 gap filling techniques for cloud removal and data reconstruction

Cloud cover creates frequent data gaps in high-resolution satellite imagery, particularly from Sentinel-2. These disrupt its continuity and reliability for time-sensitive applications such as water resource management, irrigation scheduling and crop health or yield prediction. Several approaches have been proposed in the literature, but there is a lack of performance comparisons. To address this challenge, this study aims to establish an evaluation framework to investigate the effectiveness of spatial, temporal, spatio-temporal, and spatio-spectral gap-filling approaches for restoring cloud-induced gaps in Sentinel-2 imagery using only satellite derived data. The evaluation was conducted using simulated cloud scenarios for both spatial within a single image and temporal for a time series cube of images. The methods were tested on key Sentinel-2 bands visible (B02, B03, B04), near-infrared (B08), and shortwave infrared (B11 and B12). The performance of each method was assessed using the coefficient of determination (R²), rRMSE and bias. Among the evaluated methods, spatio-temporal category, especially clustering and linear regression (CLR), showed the highest accuracy and robustness across all gap scenarios and types. Spatio-temporal Deep learning (DL) also performed well but required more training effort and failed in generalizing over all scenarios. Also, spatio-spectral approaches like SSRF showed strong results in visible and NIR bands. In contrast, spatial methods such as kriging struggled with larger or irregular gaps. These findings offer a comparative evaluation of gap-filling methods, highlighting the strengths and limitations of each approach to guide the selection of appropriate methods and explore new prospects for improving gap-filling techniques. To ensure reproducibility all codes used in this study are publicly available on https://github.com/said-grich/gap-filling .