Cloud Detection Techniques in Remote Sensing

The reliable identification of clouds and cloud shadows in satellite imagery is essential for accurate atmospheric correction, surface reflectance retrieval and downstream applications such as land-cover classification and climate monitoring. Traditional approaches rely on spectral thresholding and band‐ratio tests to distinguish clouds from other bright surfaces, yet these methods can struggle over snow, ice or urban areas. Spatial context and morphological filtering have been incorporated to reduce false positives by exploiting the contiguous structure of cloud objects. Multi-temporal analysis compares images acquired at different dates to identify anomalous bright features attributable to clouds, although this requires cloud-free reference data. In recent years, machine learning techniques—ranging from decision trees and support vector machines to ensemble methods—have offered greater robustness by learning complex spectral–spatial patterns. More recently, deep learning architectures built on convolutional neural networks have delivered pixel-level segmentation with high accuracy, even when limited to a few spectral bands, and have been optimised for rapid processing of large datasets. Active learning frameworks have further enhanced performance by iteratively incorporating human-verified samples to refine classifier boundaries. Collectively, these advances support global-scale cloud masking and facilitate near-real-time applications in environmental monitoring and disaster response.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Deep learning models designed for cloud detection in Landsat 8 imagery have achieved state-of-the-art performance. One approach employs a U-net convolutional network trained on established datasets to segment cloud pixels using only RGB bands, demonstrating improved accuracy over classical algorithms while maintaining processing times suitable for operational workflows. Another study introduced an active learning cloud detection framework for Sentinel-2 data, which generates high-quality reference masks through supervised sampling of uncertain pixels. This method supports rigorous validation of operational cloud masks produced by multiple processors and achieves overall accuracies exceeding 98%. In parallel, ready-to-use classification schemes for Sentinel-2 Multi-Spectral Imager data have been developed, utilising decision trees, Bayesian classifiers and ensemble techniques to label clouds, cirrus, shadows and clear‐sky pixels with up to 98% accuracy. These scalable algorithms provide a computationally efficient baseline for processing the ever-growing volume of Earth observation data.

Cloud Detection Techniques in Remote Sensing publication trend

The graph below shows the total number of articles in cloud detection techniques in remote sensing across all publications each year (not limited to Nature Index journals).

Technical terms

Remote sensing: Acquisition of information about the Earth’s surface via satellite or airborne sensors.

Cloud mask: A binary map indicating the presence or absence of cloud cover in each pixel of an image.

U-net architecture: A convolutional neural network design optimised for high-resolution image segmentation tasks.

Active learning: A machine learning paradigm that iteratively selects the most informative samples for labelling to improve classifier performance.

Multi-temporal analysis: The use of imagery acquired at different times to enhance detection and reduce misclassification.

Top-of-atmosphere reflectance: Spectral measurements corrected to account for solar illumination and view geometry at the sensor’s altitude.

Spectral signature: The pattern of reflectance or radiance values recorded across multiple wavelength bands.

Bayesian classifier: A statistical method that assigns class probabilities based on prior distributions and likelihoods according to Bayes’ theorem.

References

1. A cloud detection algorithm for satellite imagery based on deep learning. Remote Sensing of Environment (2019).
2. Ready-to-Use Methods for the Detection of Clouds, Cirrus, Snow, Shadow, Water and Clear Sky Pixels in Sentinel-2 MSI Images. Remote Sensing (2016).
3. Validation of Copernicus Sentinel-2 Cloud Masks Obtained from MAJA, Sen2Cor, and FMask Processors Using Reference Cloud Masks Generated with a Supervised Active Learning Procedure. Remote Sensing (2019).