Cloud Detection Capability Research Articles

Deep Learning Network Comparison and Validation in Landsat Imagery Cloud Detection

Cloud coverage poses a prevalent challenge in optical remote sensing image processing, significantly hindering the visibility and interpretability of ground‐level information. The application of deep learning technology in remote sensing image cloud detection has witnessed a remarkable advancement, rendering it an increasingly pervasive and potent solution to the challenge of cloud coverage. To thoroughly investigate the effectiveness of diverse deep learning architectures in cloud detection tasks, this research selected five seminal models: AlexNet, VGG16, GoogLeNet, ResNet34, and the cutting‐edge Swin Transformer, and performed a comprehensive comparative analysis of their application in identifying clouds within medium‐resolution Landsat 8 satellite imagery. During the training phase, each model distinctly demonstrated unique strengths in enhancing accuracy, refining loss functions, optimizing resource utilization, and accelerating training processes. In the test phase, the models’ performance was rigorously evaluated using overall accuracy (OA) and other quantitative metrics, providing solid empirical evidence for the variations in their performance. Furthermore, an insightful analysis of the visualization outcomes and a meticulous comparison of the nuanced differences between model predictions and actual cloud images underscored the individual models’ exceptional abilities in capturing intricate cloud details and executing precise edge processing. To alleviate the cumbersome process of pixel‐level labeling, this study adopts the effective block‐level labeling approach. This approach drastically streamlines annotation processes, ultimately translating into substantial cost savings and a marked reduction in time required for data preparation. The experimental results demonstrate that the Swin Transformer network performs particularly well in the cloud detection task of Landsat 8 images, achieving an OA of 90.73% under block‐level annotation, significantly higher than other comparison models, showcasing excellent cloud detection capabilities. Furthermore, the study also found that the size of input image blocks has a significant impact on detection results. Under the same network architecture, using 32 × 32 image blocks for cloud detection improves the OA by at least 10 percentage points compared to using 64 × 64 image blocks. This finding provides important insights for optimizing cloud detection algorithms. In summary, this study not only verifies the superiority of Swin Transformer in cloud detection in remote sensing images but also reveals the impact of image block size on detection performance, providing new ideas and directions for future advancements in remote sensing image processing and cloud detection technology.

Incorporating EarthCARE observations into a multi-lidar cloud climate record: the ATLID (Atmospheric Lidar) cloud climate product

Abstract. Despite significant advances in atmospheric measurements and modeling, clouds' response to human-induced climate warming remains the largest source of uncertainty in model predictions of climate. The launch of the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite in 2006 started the era of long-term spaceborne optical active sounding of Earth's atmosphere, which continued with the CATS (Cloud-Aerosol Transport System) lidar on board the International Space Station (ISS) in 2015 and the Atmospheric Laser Doppler Instrument (ALADIN) lidar on board Aeolus in 2018. The next important step is the Atmospheric Lidar (ATLID) instrument from the EarthCARE (Earth Clouds, Aerosols and Radiation Explorer) mission, expected to launch in 2024. In this article, we define the ATLID Climate Product, Short-Term (CLIMP-ST) and ATLID Climate Product, Long-Term (CLIMP-LT). The purpose of CLIMP-ST is to help evaluate the description of cloud processes in climate models, beyond what is already done with existing space lidar observations, thanks to ATLID's new capabilities. The CLIMP-LT product will merge the ATLID cloud observations with previous space lidar observations to build a long-term cloud lidar record useful to evaluate the cloud climate variability predicted by climate models. We start with comparing the cloud detection capabilities of ATLID and CALIOP (Cloud-Aerosol Lidar with Orthogonal Polarization) in day- and nighttime, on a profile-to-profile basis in analyzing virtual ATLID (355 nm) and CALIOP (532 nm) measurements over synthetic cirrus and stratocumulus cloud scenes. We show that solar background noise affects the cloud detectability in daytime conditions differently for ATLID and CALIPSO. We found that the simulated daytime ATLID measurements have lower noise than the simulated daytime CALIOP measurements. This allows for lowering the cloud detection thresholds for ATLID compared to CALIOP and enables ATLID to better detect optically thinner clouds than CALIOP in daytime at high horizontal resolution without false cloud detection. These lower threshold values will be used to build the CLIMP-ST (Short-Term, related only to the ATLID observational period) product. This product should provide the ability to evaluate optically thin clouds like cirrus in climate models compared to the current existing capability. We also found that ATLID and CALIPSO may detect similar clouds if we convert ATLID 355 nm profiles to 532 nm profiles and apply the same cloud detection thresholds as the ones used in GOCCP (GCM-Oriented CALIPSO Cloud Product; general circulation model). Therefore, this approach will be used to build the CLIMP-LT product. The CLIMP-LT data will be merged with the GOCCP data to get a long-term (2006–2030s) cloud climate record. Finally, we investigate the detectability of cloud changes induced by human-caused climate warming within a virtual long-term cloud monthly gridded lidar dataset over the 2008–2034 period that we obtained from two ocean–atmosphere coupled climate models coupled with a lidar simulator. We found that a long-term trend of opaque cloud cover should emerge from short-term natural climate variability after 4 years (possible lifetime) to 7 years (best-case scenario) for ATLID merged with CALIPSO measurements according to predictions from the considered climate models. We conclude that a long-term lidar cloud record built from the merging of the actual ATLID-LT data with CALIPSO-GOCCP data will be a useful tool for monitoring cloud changes and evaluating the realism of the cloud changes predicted by climate models.

Joint multiscale cloud detection algorithm for ground-based lidar.

A ground-based lidar is a powerful tool for studying the vertical structure and optical properties of clouds. A layer detection algorithm is important to determine the presence and spatial position of clouds from vast lidar signals. However, current detection algorithms for ground-based lidar still involve substantial missing and false detections for tenuous layers and layer edges. Here, a joint multiscale cloud layer detection algorithm is proposed. The algorithm can effectively capture the tenuous layers and layer edges by using joint multiscale detection methods based on a trend function and the Bernoulli distribution assumption. Results show that the proposed algorithm detects 10.45% more cloud layers than the official cloud product of Micro Pulse Lidar Network (MPLNET) does. Specifically, 7.93% and 12.57% more cloud layers are detected at daytime and nighttime, respectively. The evaluation based on depolarization properties proves that the additional cloud layers detected by the joint multiscale algorithm are reliable. These additional detected clouds have important implications for cloud climatology and climate change research. The new algorithm remarkably enhances the cloud detection capability of ground-based lidar and potentially be widely used by the community.

Cloud detection algorithm for multi-modal satellite imagery using convolutional neural-networks (CNN)

Cloud detection algorithms are crucial in many remote-sensing applications to allow an optimized processing of the acquired data, without the interference of the cloud fields above the surfaces of interest (e.g., land, coral reefs, etc.). While this is a well-established area of research, replete with a number of cloud detection methodologies, many issues persist for detecting clouds over areas of high albedo surfaces (snow and sand), detecting cloud shadows, and transferring a given algorithm between observational platforms. Particularly for the latter, algorithms are often platform-specific with corresponding rule-based tests and thresholds inherent to instruments and applied corrections. Here, we present a convolutional neural network (CNN) algorithm for the detection of cloud and cloud shadow fields in multi-channel satellite imagery from World-View-2 (WV-2) and Sentinel-2 (S-2), using their Red, Green, Blue, and Near-Infrared (RGB, NIR) channels. This algorithm is developed within the NASA NeMO-Net project, a multi-modal CNN for global coral reef classification which utilizes imagery from multiple remote sensing aircraft and satellites with heterogeneous spatial resolution and spectral coverage. Our cloud detection algorithm is novel in that it attempts to learn deep invariant features for cloud detection utilizing both the spectral and the spatial information inherent in satellite imagery. The first part of our work presents the CNN cloud and cloud shadow algorithm development (trained using WV-2 data) and its application to WV-2 (with a cloud detection accuracy of 89%) and to S-2 imagery (referred to as augmented CNN). The second part presents a new domain adaptation CNN-based approach (domain adversarial NN) that allows for better adaptation between the two satellite platforms during the prediction step, without the need to train for each platform separately. Our augmented CNN algorithm results in better cloud prediction rates as compared to the original S-2 cloud mask (81% versus 48%), but still, clear pixels prediction rate is lower than S-2 (81% versus 91%). Nevertheless, the application of the domain adaptation approach shows promise in better transferring the knowledge gained from one trained domain (WV-2) to another (S-2), increasing the prediction accuracy of both clear and cloudy pixels when compared to a network trained only by WV-2. As such, domain adaptation may offer a novel means of additional augmentation for our CNN-based cloud detection algorithm, increasing robustness towards predictions from multiple remote sensing platforms. The approach presented here may be further developed and optimized for global and multi-modal (multi-channel and multi-platform) satellite cloud detection capability by utilizing a more global dataset.

Data Assimilation of Geostationary Microwave Simulation for Hurricane Sandy

Geostationary orbit microwave atmospheric sounding based on interferometric technology is the newest field of microwave remote sensing. Equipped with a microwave atmospheric remote sensing instruments on the geostationary orbit meteorological satellite platform will increase the frequency of observations at the same time, improve the cloud detection capability. The current polar-orbiting meteorological satellite observation system ensures that only six hours of observation period, unable to typhoons and other fast-changing weather systems for real-time observation. This paper mainly focuses on the following two parts: 1) data assimilation of geostationary microwave observations with high temporal and spatial resolution for the full disk and 2) Typhoons path predicted and heavy precipitation inversion.

Investigating the frequency and interannual variability in global above-cloud aerosol characteristics with CALIOP and OMI

Abstract. Seven and a half years (June 2006 to November 2013) of Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) aerosol and cloud layer products are compared with collocated Ozone Monitoring Instrument (OMI) aerosol index (AI) data and Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) cloud products in order to investigate variability in estimates of biannual and monthly above-cloud aerosol (ACA) events globally. The active- (CALIOP) and passive-based (OMI-MODIS) techniques have their advantages and caveats for ACA detection, and thus both are used to derive a thorough and robust comparison of daytime cloudy-sky ACA distribution and climatology. For the first time, baseline above-cloud aerosol optical depth (ACAOD) and AI thresholds are derived and examined (AI = 1.0, ACAOD = 0.015) for each sensor. Both OMI-MODIS and CALIOP-based daytime spatial distributions of ACA events show similar patterns during both study periods (December–May) and (June–November). Divergence exists in some regions, however, such as Southeast Asia during June through November, where daytime cloudy-sky ACA frequencies of up to 10 % are found from CALIOP yet are non-existent from the OMI-based method. Conversely, annual cloudy-sky ACA frequencies of 20–30 % are reported over northern Africa from the OMI-based method yet are largely undetected by the CALIOP-based method. Using a collocated OMI-MODIS-CALIOP data set, our study suggests that the cloudy-sky ACA frequency differences between the OMI-MODIS- and CALIOP-based methods are mostly due to differences in cloud detection capability between MODIS and CALIOP as well as QA flags used. An increasing interannual variability of ∼ 0.3–0.4 % per year (since 2009) in global monthly cloudy-sky ACA daytime frequency of occurrence is found using the OMI-MODIS-based method. Yet, CALIOP-based global daytime ACA frequencies exhibit a near-zero interannual variability. Further analysis suggests that the OMI-derived interannual variability in cloudy-sky ACA frequency may be affected by OMI row anomalies in later years. A few regions are found to have increasing slopes in interannual variability in cloudy-sky ACA frequency, including the Middle East and India. Regions with slightly negative slopes of the interannual variability in cloudy-sky ACA frequencies are found over South America and China, while remaining regions in the study show nearly zero change in ACA frequencies over time. The interannual variability in ACA frequency is not, however, statistically significant on both global and regional scales, given the relatively limited sample sizes. A longer data record of ACA events is needed in order to establish significant trends of ACA frequency regionally and globally.

Enhancing a Simple MODIS Cloud Mask Algorithm for the Landsat Data Continuity Mission

The upcoming Landsat Data Continuity Mission (LDCM) will include new channels centered around 1.38 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\mu\hbox{m}$</tex></formula> and 12 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\mu \hbox{m}$</tex></formula> . This work studies the potential impact of these new channels on LDCM's cloud detection capabilities by using MODerate resolution Imaging Spectroradiometer (MODIS) data as a proxy. Thresholds for the 1.38 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\mu\hbox{m}$</tex></formula> band and the so-called “split window” technique (using the brightness temperature difference of bands centered at 11 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$ \mu\hbox{m}$</tex></formula> and 12 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\mu\hbox{m}$</tex></formula> ) are derived using atmospheric profiles from the ECMWF ERA-40 reanalysis and a MODIS-band radiance simulator. The thresholds are incorporated into a previously published cloud mask scheme and applied on low- and mid-latitude (60 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$^{\circ}$</tex></formula> S to 60 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$^{\circ}\ \hbox{N}$</tex></formula> ) MODIS radiance data from two different days, six months apart. While the previous scheme yields agreement rates to the MODIS cloud mask just below 80 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\%$</tex></formula> , the addition of the 1.38 <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$\mu\hbox{m}$</tex></formula> and split window tests increases the agreement by 7–9%. The earlier scheme is still appropriate for cloud masking of historical Landsat images and for carrying consistent cloud detection into the future. The enhanced scheme of this paper, on the other hand, with its improved masking of primarily high thin clouds, can be either used independently or combined with other masking techniques for generating reliable LDCM cloud mask products that can potentially include confidence indicators based on the degree of agreement between multiple cloud masks.

Potential of 229-GHz Channel for Ice-Cloud Screening

The purpose of this paper is to evaluate the potential of a 229-GHz channel as part of a humidity sounder to improve the sounder's cloud-screening capability. The humidity-sounding channels centered around the 183.31-GHz water-vapor line are affected by ice-cloud scattering. Usually, in the operational assimilation of humidity data, the problem of undetected clouds is overcome by setting high observation errors for humidity channels. However, high observation errors diminish the impact of good clear-sky data also, thus, leading to a smaller impact from humidity sounders in weather forecasts. It is expected that a channel at 229 GHz can offer a better cloud-detection capability due to its higher sensitivity to ice scattering compared with the existing channels. Here, a simple cloud-screening method is devised based on the difference between observed and background brightness temperatures. This method was tested for 229 GHz and for an existing channel at 166 GHz, and the results are compared. The new method using 229 GHz is also compared with an existing cloud-screening test, namely, the cirrus-cost test which uses the 183-GHz channels. The study demonstrates that a channel at 229 GHz has a better skill at detecting cirrus clouds which could be interpreted as an error in the background humidity profile. Therefore, addition of a 229-GHz channel can allow more weight to be given to the 183-GHz channels in data assimilation thus significantly increasing their impact on weather forecasting.

Comparison of macroscopic cloud data from ground-based measurements using VIS/NIR and IR instruments at Lindenberg, Germany

A comparison between different types of ground-based sensors has been carried out to derive macroscopic cloud data such as cloud cover and cloud-base heights. The instruments compared in the campaign at the Meteorological Observatory Lindenberg in the period May to September 2006 include an infrared (IR) sky scanner called Nubiscope, a Daylight VIS/NIR Whole Sky Imager (WSI), a ceilometer LD-40 measuring in the near infrared region (NIR) and a Ka band cloud radar measuring in the micro wave band (extremely high frequency or EHF) region. In addition, our data analysis included regular hourly cloud observations by weather observers, and vertical profiles of temperature, humidity and winds taken from six-hourly radio soundings at the site. The comparison has been focused on performance and features of the Nubiscope as a prototype instrument for automatic cloud observations. Cloud cover (CC) derived from the Nubiscope cloud algorithm compares quite well with CC derived from both WSI and from observations. CC differences are within ± 2 Okta in 67% of cases between Nubiscope and observations, and in 90% of cases between Nubiscope and WSI. The cloud detection capability as derived from the zenith signals of Nubiscope and WSI shows coincidence in about 90% of cases. For cloud-base heights (CBHs) from Nubiscope data and ceilometer as well as from radar reflectivity, the comparison showed a general good correspondence in the lower and middle troposphere up to heights of about 6 km with some systematic difference due to the different detection methods. For the upper troposphere above 6 km the differences become widespread and more random. Cloud detection capabilities of the instruments are also illustrated by a case study of moving clouds with patterns similar to contrails that were erroneously classified as such by the weather observer mainly due to lack of height information that the ceilometer did not provide. By combined information from WSI, radio sonde humidity and radar, they were shown not to be contrails, but most likely low-level water clouds either of natural origin or built from aircraft at their ascent or descent flight close to the airport.

A comparison of MODIS-derived cloud amount with visual surface observations

Two main sources for global cloud climatologies are visual surface observations and observations made by spaceborne sensors. Satellite observations compared with surface data show in most cases differences ranging from − 15% up to − 1%, depending on sensor and observation conditions. These differences are partially controlled by sensors' cloud detection capabilities — a higher number of spectral bands and higher spatial resolution are believed to allow discrimination of clouds from land/ocean/snow background. A Moderate-Resolution Imaging Spectroradiometer (MODIS) produces images of the atmosphere in 36 spectral bands with a spatial resolution of 250–1000 m, thus having a capacity for cloud detection far more advanced than other operating sensors. In this study, instantaneous MODIS cloud observations were compared with surface data for Poland for January (winter) and July (summer) 2004. It was found that MODIS observed 4.38% greater cloud amount in summer conditions and 7.28% in winter conditions. Differences were greater at night (7–8%) than in daytime (0.5–7%) and correlations ranged between 0.577 (winter night) and 0.843 (winter day, summer day and night).

Reconfigurable processing for satellite on-board automatic cloud cover assessment

Clouds have a critical role in many studies such as weather- and climate-related investigations. However, they represent a source of errors in many applications, and the presence of cloud contamination can hinder the use of satellite data. In addition, sending cloudy data to ground stations can result in an inefficient utilization of the communication bandwidth. This requires satellite on-board cloud detection capability to mask out cloudy pixels from further processing. Remote sensing satellite missions have always required smaller size, lower cost, more flexibility, and higher computational power. Reconfigurable Computers (RCs) combine the flexibility of traditional microprocessors with the power of Field Programmable Gate Arrays (FPGAs). Therefore, RCs are a promising candidate for on-board preprocessing. This paper presents the design and implementation of an RC-based real-time cloud detection system. We investigate the potential of using RCs for on-board preprocessing by prototyping the Landsat 7 ETM+ ACCA algorithm on one of the state-of-the-art reconfigurable platforms, SRC-6. It will be shown that our work provides higher detection accuracy and over one order of magnitude improvement in performance when compared to previously reported investigations.

Relative merits of the 1.6 and 3.75 μm channels of the AVHRR/3 for cloud detection

A study was performed to investigate the potential impacts of the cloud-masking capability of the advanced very high resolution radiometer (AVHRR) on board the National Oceanic and Atmospheric Administration (NOAA) polar orbiting satellites because of the addition of the 1.6 µm channel (ch3a) and the removal of the 3.75 µm channel (ch3b) during daylight operation. Both channels are measured by the AVHRR, but only one is available in the data stream. Because the AVHRR presents the longest time series of global imager data, this change could impact a critical source of climate data. Specifically, changes in its cloud-masking capability may introduce a discontinuity in its derived clear-sky data records. To study the relative cloud-detection capabilities of the AVHRR with ch3a or ch3b, data from the moderate resolution imaging spectroradiometer (MODIS) on the National Aeronautics and Space Administration (NASA) TERRA mission was used because it offers ch3a, ch3b, and the other AVHRR channels simultaneously. The MODIS data analysis indicated that ch3b offered more capability in separating cloud from snow. Cloud-masking results with ch3a and ch3b were comparable with respect to the separation of cloud from aerosol and the detection of cloud over a desert scene. Although these results are preliminary and based on a limited analysis, they do indicate that the switch from ch3b to ch3a may have significant impacts on the cloud-detection capability of the AVHRR.

Detection of Nonprecipitating Clouds with the WSR-88D: A Theoretical and Experimental Survey of Capabilities and Limitations

Abstract Theoretical calculations and experiments verify that the National Weather Service WSR-88D radars have the sensitivity to detect nonprecipitating clouds, but show that significant obstacles impair the generality of this cloud sensing technique. Bragg scatter from refractive index inhomogeneities can be of the same magnitude as cloud echoes under many conditions, whereupon interpretation of WSR-88D echoes can be either complicated or impossible. Moreover, problems with echoes from ground clutter and from insects, birds, and other floating debris hinder WSR-88D cloud detection capabilities, particularly at low elevation angles. To illustrate these problems WSR-88D reflectivities collected using volume coverage pattern (VCP) 21 from the months of March and October 1996 were compared with collocated reflectivities obtained from a zenith-pointing 94-GHz cloud radar located in central Pennsylvania. Coincident echo detection occurred 82% and 39% of the time, while the WSR-88D had significant detections 7...

Simulation of satellite lidar and radiometer retrievals of a general circulation model three‐dimensional cloud data set

The inclusion of a backscatter lidar on a space platform for a radiation mission, as proposed by various space agencies, aims to bring new information on three‐dimensional cloud distribution, with a special emphasis on optically thin cirrus clouds, which are presently poorly detected by passive sensors. Key issues for such cloud observational studies are the detection of multilayered cloud systems, thin cirrus, and fractional cloud cover, knowledge that would improve our understanding of the global radiation budget. To assess the impact of such lidar measurements on cloud climatology, a 1 month cloud data set has been simulated with a general circulation model (GCM). The cloud detection capability of a spaceborne scanning backscatter lidar is assessed with the use of two detection schemes, one based on limitations in the detected cloud optical depth and the other based on lidar signal‐to‐noise ratio. The cloud information retrieved from passive radiometric measurements using a procedure like that used in the International Satellite Cloud Climatology Project is also simulated from the same GCM cloud data set. It is shown that a spaceborne backscatter lidar can improve significantly the retrieval of thin cirrus clouds as well as underlying cloud layers. High‐level cloud retrieval from a spaceborne lidar therefore appears as a powerful complement to radiometric measurements for improving our knowledge of actual cloud climatology.