High-Quality Cloud Masking of Landsat 8 Imagery Using Convolutional Neural Networks

Hughes, Michael J.; Kennedy, Robert J.

doi:10.3390/rs11212591

Cited by 38 publications

(30 citation statements)

References 32 publications

(22 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…They validated their models on images of the same dataset and the opposite one (models trained on L8-Biome were tested on L8-SPARCS and vice versa), showing significant improvements over FMask, especially for snow/ice covers. Hughes and Kennedy [41] also trained and validated a U-Net using different images of the L8-SPARCS dataset, claiming an accuracy on par with human interpreters. Wieland et al [40] also trained a U-Net on the L8-SPARCS dataset, which was validated on a custom private dataset of 14 1024 × 1024 Landsat-7, Landsat-8, and Sentinel-2 images.…”

Section: Related Work On Deep Learning Models For Cloud Detectionmentioning

confidence: 99%

“…Table 2 shows some of the most recent works on deep learning for cloud masking, applied to Landsat-8 and Sentinel-2 imagery, depending on the validation scheme used. The predominance of the intra-dataset scheme is notorious; The intra-dataset scheme tends to lead to a very high cloud detection accuracy and larger gains when compared with threshold-based methods [8,35,41]; however, in several cases, these models are trained on relatively small or local datasets (L8-SPARCS or L8-38Clouds), which might not be representative of global land cover and cloud conditions. One strategy to determine whether these models generalize to different conditions is to test them over independent datasets following an inter-dataset strategy.…”

Section: Validation Approachesmentioning

confidence: 99%

“…Finally, very few works address cross-sensor validation [15,40] and none of them for public Sentinel-2 data, which is one of the contributions of our work. [41] SegNet [11] Cloud-Net [20] MSCFF [8] CloudX-net [13] CDnet [12] Lightweight U-Net [14] LeNet […”

Section: Validation Approachesmentioning

confidence: 99%

See 2 more Smart Citations

Benchmarking Deep Learning Models for Cloud Detection in Landsat-8 and Sentinel-2 Images

2021

View full text Add to dashboard Cite

The systematic monitoring of the Earth using optical satellites is limited by the presence of clouds. Accurately detecting these clouds is necessary to exploit satellite image archives in remote sensing applications. Despite many developments, cloud detection remains an unsolved problem with room for improvement, especially over bright surfaces and thin clouds. Recently, advances in cloud masking using deep learning have shown significant boosts in cloud detection accuracy. However, these works are validated in heterogeneous manners, and the comparison with operational threshold-based schemes is not consistent among many of them. In this work, we systematically compare deep learning models trained on Landsat-8 images on different Landsat-8 and Sentinel-2 publicly available datasets. Overall, we show that deep learning models exhibit a high detection accuracy when trained and tested on independent images from the same Landsat-8 dataset (intra-dataset validation), outperforming operational algorithms. However, the performance of deep learning models is similar to operational threshold-based ones when they are tested on different datasets of Landsat-8 images (inter-dataset validation) or datasets from a different sensor with similar radiometric characteristics such as Sentinel-2 (cross-sensor validation). The results suggest that (i) the development of cloud detection methods for new satellites can be based on deep learning models trained on data from similar sensors and (ii) there is a strong dependence of deep learning models on the dataset used for training and testing, which highlights the necessity of standardized datasets and procedures for benchmarking cloud detection models in the future.

show abstract

Section: Related Work On Deep Learning Models For Cloud Detectionmentioning

confidence: 99%

Section: Validation Approachesmentioning

confidence: 99%

See 1 more Smart Citation

Benchmarking Deep Learning Models for Cloud Detection in Landsat-8 and Sentinel-2 Images

2021

View full text Add to dashboard Cite

show abstract

“…From the generated 10 sub-samples we calculated the following metrics: overall accuracy, Cohen's Kappa coefficient, and F1-score. These metrics were selected as they are widely used to evaluate classification methods [71,72]. We further evaluated the methods using the precision-recall ratio [73] which provides insights on the trade-off between precision (error of commission) and recall (error of omission) rates for each method, where a value of 1 is preferred means there the precision and recall scores are equal and balanced.…”

Section: Evaluation Designmentioning

confidence: 99%

Comparing Sentinel-1 Surface Water Mapping Algorithms and Radiometric Terrain Correction Processing in Southeast Asia Utilizing Google Earth Engine

Markert

Mayer

et al. 2020

Remote Sensing

View full text Add to dashboard Cite

Satellite remote sensing plays an important role in the monitoring of surface water for historical analysis and near real-time applications. Due to its cloud penetrating capability, many studies have focused on providing efficient and high quality methods for surface water mapping using Synthetic Aperture Radar (SAR). However, few studies have explored the effects of SAR pre-processing steps used and the subsequent results as inputs into surface water mapping algorithms. This study leverages the Google Earth Engine to compare two unsupervised histogram-based thresholding surface water mapping algorithms utilizing two distinct pre-processed Sentinel-1 SAR datasets, specifically one with and one without terrain correction. The resulting surface water maps from the four different collections were validated with user-interpreted samples from high-resolution Planet Scope data. It was found that the overall accuracy from the four collections ranged from 92% to 95% with Cohen’s Kappa coefficients ranging from 0.7999 to 0.8427. The thresholding algorithm that samples a histogram based on water edge information performed best with a maximum accuracy of 95%. While the accuracies varied between methods it was found that there is no statistical significant difference between the errors of the different collections. Furthermore, the surface water maps generated from the terrain corrected data resulted in a intersection over union metrics of 95.8%–96.4%, showing greater spatial agreement, as compared to 92.3%–93.1% intersection over union using the non-terrain corrected data. Overall, it was found that algorithms using terrain correction yield higher overall accuracy and yielded a greater spatial agreement between methods. However, differences between the approaches presented in this paper were not found to be significant suggesting both methods are valid for generating accurate surface water maps. High accuracy surface water maps are critical to disaster planning and response efforts, thus results from this study can help inform SAR data users on the pre-processing steps needed and its effects as inputs on algorithms for surface water mapping applications.

show abstract

“…The main methods include fully connected deep neural networks (FCDNs) and convolution neural networks (CNNs) [21]. Applying the deep learning method for CSM has been explored in the recent decade [22][23][24]. Most of these researches used CNN to identify cloud or clear-sky features from the satellite images in a regional area, and the model accuracy and efficiency were not validated globally.…”

Section: Introductionmentioning

confidence: 99%

A Deep Learning Trained Clear-Sky Mask Algorithm for VIIRS Radiometric Bias Assessment

et al. 2019

View full text Add to dashboard Cite

Clear-sky mask (CSM) is a crucial influence on the calculating accuracy of the sensor radiometric biases for spectral bands of visible, infrared, and microwave regions. In this study, a fully connected deep neural network (FCDN) was proposed to generate CSM for the Visible Infrared Imaging Radiometer Suite (VIIRS) onboard Suomi National Polar-Orbiting Partnership (S-NPP) and NOAA-20 satellites. The model, well-trained by S-NPP data, was used to generate both S-NPP and NOAA-20 CSMs for the independent data, and the results were validated against the biases between the sensor observations and Community Radiative Transfer Model (CRTM) calculations (O-M). The preliminary result shows that the FCDN-CSM model works well for identifying clear-sky pixels. Both O-M mean biases and standard deviations were comparable with the Advance Clear-Sky Processor over Ocean (ACSPO) and were significantly better than a prototype cloud mask (PCM) and the case without a clear-sky check. In addition, by replacing CRTM brightness temperatures (BTs) with the atmosphere air temperature and water vapor contents as input features, the FCDN-CSM exhibits its potential to generate fast and accurate VIIRS CSM onboard follow-up Joint Polar Satellite System (JPSS) satellites for sensor calibration and validation before the physics-based CSM is available.

show abstract

High-Quality Cloud Masking of Landsat 8 Imagery Using Convolutional Neural Networks

Cited by 38 publications

References 32 publications

Benchmarking Deep Learning Models for Cloud Detection in Landsat-8 and Sentinel-2 Images

Benchmarking Deep Learning Models for Cloud Detection in Landsat-8 and Sentinel-2 Images

Comparing Sentinel-1 Surface Water Mapping Algorithms and Radiometric Terrain Correction Processing in Southeast Asia Utilizing Google Earth Engine

A Deep Learning Trained Clear-Sky Mask Algorithm for VIIRS Radiometric Bias Assessment

Contact Info

Product

Resources

About