.png){kind=icon}
.png){kind=icon}
Return
Make more pixels usable: A data-driven framework for customizable cloud and shadow processing in satellite imagery
.png){kind=icon}
PRE
AI
DOI：10.1016/j.isprsjprs.2026.04.039.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Atmospheric interference, particularly cloud cover, remains a persistent challenge in the effective utilization of optical remote sensing imagery. Existing cloud processing methods either localize cloud positions or remove cloud interference but fail to adaptively process cloudy imagery, leading to substantial loss of valuable pixel information. To address this limitation, this paper introduces a data-driven framework for optimizing pixel utilization in satellite images. First, we proposed an adjustable cloud intensity index to quantify the interference of cloud and shadow, facilitating selective processing of different cloud pixels. Furthermore, a novel cloud simulation strategy was applied to Landsat-8 data, generating globally-distributed cloud samples and constructing the benchmark CloudL8-LC dataset. The CloudL8-LC dataset comprises over 5.2 billion pixels and has the widest coverage among the existing cloud datasets, and supports cloud detection, cloud removal, and adjustable cloud processing. Based on this dataset, we developed a single-image cloud intensity estimation network (SCIENet) equipped with a relevance-aware attention module (RAAM) and a gating-based channel-spatial feature pruner (GCSFP), both designed to accommodate the inherent variability of cloud formations. Once the cloud intensity index is predicted by SCIENet, the threshold can be flexibly adjusted to accommodate user-specific requirements, enabling customizable cloud and shadow processing within single imagery. Additionally, we introduced a suite of evaluation metrics to assess cloud-processing performance in terms of pixel quantity, quality, and practical utility. Extensive experiments demonstrated that the proposed cloud intensity index functions as a dynamic thresholding parameter for single-image cloud processing, enabling selective masking of highly contaminated pixels while reconstructing minimally affected regions. This provides users with tunable cloud processing solutions and exhibits robustness across diverse cloud types. Compared to cloud detection approaches, SCIENet significantly enhances pixel availability while maintaining reliable classification accuracy, ensuring consistent performance across various land-cover scenarios. Furthermore, in some real-world cloud conditions where the classic cFmask algorithm achieves a cloud detection precision below 70%, the proposed SCIENet model attains a precision of up to 90%. Additionally, relative to manually annotated masks, the SCIENet can increase the number of clear pixels by over 15%, making more pixels usable.
Keywords:
cloud processing
satellite imagery
cloud intensity index
pixel utilization
data-driven framework
Journal
IF:
12.2
Papers:
4.3K
Citations:
3.2W
View more organizations