Cloud Detection Algorithm Research Articles

Swincloud: a hybrid network for cloud detection in thermal infrared remote sensing images

ABSTRACT Cloud cover is a significant factor affecting the effectiveness of satellite-based Earth observations. Existing cloud detection algorithms primarily rely on imaging data from satellite sensors in the visible to near-infrared spectral range, making it challenging to achieve day-and-night cloud monitoring. Convolutional neural networks have shown outstanding performance in previous cloud detection algorithms due to their robust ability to extract local information. However, their inherent inductive bias limits their capacity to learn long-range semantic information. To address these challenges, we proposed SwinCloud, a U-shaped semantic segmentation network based on an enhanced Swin Transformer for cloud detection in the thermal infrared spectral range. Specifically, we augment the Swin Transformer's window attention module with a CNN-based parallel pathway to effectively model global-local information. We employ a feature fusion module before the final upsampling module in the decoder to better integrate low-level spatial information and high-level semantic information. On the Landsat-8 cloud detection dataset, our model outperforms state-of-the-art methods. When transferred to the SDGSAT-TIS cloud detection dataset, the mIOU of experiment results reaches 69.9%, demonstrating the strong transferability of SwinCloud across different sensors. We also applied SwinCloud to cloud detection in the visible bands of Landsat-8. The results demonstrated SwinCloud's generalization capability across different bands.

A method to assess the cloud-aerosol transition zone from ceilometer measurements

Cloud and aerosol contribution to the Earth's radiative budget constitutes one of the most significant uncertainties in future climate projections. To distinguish between clouds and cloud-free air, atmospheric scientists have been using a wide range of instruments, techniques, and algorithms with various detection thresholds. However, since the change from a cloudy to a cloud-free atmosphere may be gradual and, in some cases, far from obvious, recent research is questioning where the threshold between both phases should be established. These considerations lead to contemplate a transition zone (TZ) between cloud and cloud-free conditions. In the present study, backscatter profiles obtained by a Vaisala CL31 ceilometer were processed to assess the transition zone. First, two widely used cloud detection algorithms were applied and compared: the method provided by the ceilometer manufacturer (Vaisala) and Cloudnetpy, the algorithm from ACTRIS Cloudnet, a project devoted to aerosol, clouds, and trace gases research. Second, a sensitivity analysis was applied to the backscatter and signal-to-noise ratio thresholds used for cloud detection in Cloudnetpy. This methodology has allowed us to assess the vertical distribution of clouds, aerosols, the TZ, and its frequency of occurrence. Results indicate a gradual transition in backscatter retrievals from cloud to cloud-free, where particles detected near cloud boundaries induced higher backscatter values than those found further away. Depending on the thresholds used, we observed a 9.3% (with an estimated range of uncertainty of 5.420%) variation in cloud occurrence which can be attributed to TZ conditions. Analysing the whole backscatter profile, we found as many TZ conditions as cloudy values, which emphasises the importance of studying the vertical distribution of the TZ. Moreover, the analysis of TZ occurrence in height and time revealed that such conditions concentrate below 800m during night periods, although annual height-hour distributions involve a remarkable variability among seasons. These findings highlight the importance of either including an additional phase between ‘pure clouds’ and ‘pure aerosols’ or treating them as a continuum of suspended particles in the atmosphere.

A novel spectro-temporal index NDTeI combined with machine learning algorithm for Sentinel-2 cloud detection

ABSTRACT There are challenges for cloud detection over bright surfaces, especially for Sentinel-2 image without thermal band. In this paper, we developed a new cloud detection algorithm for Sentinel-2 data designed to tackle the challenges of cloud detection on bright targets. Our algorithm (called RF-NDTeI-SNIC) was built through the integration of Random Forest (RF) model, Normalized Difference Temporal Index (NDTeI) cloud index and Simple Non-Iterative Clustering (SNIC) image segmentation algorithm. First, RF and NDTeI were used to derive the initial and second stage’s cloud detection results, respectively. Then, the RF- and NDTeI-based results were fused to remove the bright surfaces and land cover changes. Finally, the fusion results were further refined using SNIC for morphological processing. The proposed RF-NDTeI-SNIC algorithm was evaluated using nine challenging cloud-covered Sentinel-2 scenes with substantial bright surfaces. Results indicated that the average overall accuracy of our algorithm was 95.01%, with average commission rate of 10.19% for cloud. In addition, five commonly used Sentinel-2 cloud mask algorithms including s2cloudless, QA, Fmask, CDI and Tmask were selected for comparative analysis. Results suggested that our algorithm outperformed others with overall accuracies of 89.60–94.17% and cloud commission rate of 9.46–37.50%, and have significant advantages in terms of bright surfaces. In summary, the RF-NDTeI-SNIC algorithm we developed was capable of yielding accurate Sentinel-2 cloud masks, and the novel NDTeI cloud index we proposed gave a new and effective approach for improving the separability of clouds and bright surfaces.

Validation of cloud mask product from multi-channel satellite INSAT-3DR

The Indian National Satellite-3DR (INSAT-3DR) provides half-hourly multispectral images over the Indian region from a geostationary position. It observes diurnal variations highly relevant for variety of application areas, including weather forecasting. This work assesses the performance of the cloud mask product of INSAT-3DR new algorithm developed by Indian Space Research Organization (ISRO) by comparison with Moderate Resolution Imaging Spectrometer radiometer (MODIS) cloud mask data for the months from July 2019 to July 2020. The cloud detection algorithm employs nine different tests, considering solar illumination, satellite angle and surface type conditions to get pixel-resolution cloud mask. Our validation results indicate that the probability of detection (POD) of cloudy (clear) sky is 86 % (82 %) with 83 % hit rate for INSAT-3DR cloud mask, using MODIS cloud mask as a reference. When the INSAT-3DR cloud mask algorithm is also implemented on similar channels of MODIS, the POD of cloudy (clear) sky is 76 % (74 %) with 85 % hit rate. Our findings indicate the reliability of the INSAT-3DR cloud mask retrieved product for research and other applications. This is one of the important parameter of satellite derived products, which plays major role in daily operational weather forecasting.

Advancements in Satellite Remote Sensing for Predicting Large-Scale Wildfire Risks: An Image Processing Algorithmic Framework

With the escalation of global warming and human activities, large-scale wildfires have become increasingly frequent, posing significant threats to both ecological environments and human societal safety. Satellite remote sensing technology plays a pivotal role in wildfire monitoring and risk assessment, providing extensive geographical coverage and continuous monitoring capabilities. Traditional methods for predicating wildfire risk, however, face limitations in processing large-scale remote sensing data, especially in cloud detection and temporal information analysis. In response to this challenge, a novel set of image processing algorithms has been developed to enhance the efficiency and accuracy of wildfire risk prediction. Initially, a cloud detection and removal algorithm based on deep learning is introduced. This algorithm effectively identifies and eliminates cloud interference in remote sensing images, thereby significantly improving the quality and usability of image data. Subsequently, a temporal information capturing technique is proposed, capable of processing vast amounts of remote sensing data and extracting time series features. This technique provides robust data support for wildfire risk prediction. The application of these technologies not only improves the efficiency of the data processing workflow but also enhances the timeliness and accuracy of the prediction model, holding significant practical importance for guiding actual wildfire prevention and response measures.

Cloud properties across spatial scales in simulations of the interstellar medium

Molecular clouds (MCs) are structures of dense gas in the interstellar medium (ISM) that extend from ten to a few hundred parsecs and form the main gas reservoir available for star formation. Hydrodynamical simulations of a varying complexity are a promising way to investigate MCs evolution and their properties. However, each simulation typically has a limited range in resolution and different cloud extraction algorithms are used, which complicates the comparison between simulations. In this work, we aim to extract clouds from different simulations covering a wide range of spatial scales. We compare their properties, such as size, shape, mass, internal velocity dispersion, and virial state. We applied the Hop cloud detection algorithm on (M)HD numerical simulations of stratified ISM boxes and isolated galactic disk simulations that were produced using Flash Ramses and Arepo We find that the extracted clouds are complex in shape, ranging from round objects to complex filamentary networks in all setups. Despite the wide range of scales, resolution, and sub-grid physics, we observe surprisingly robust trends in the investigated metrics. The mass spectrum matches in the overlap between simulations without rescaling and with a high-mass power-law index of -1 for logarithmic bins of mass, in accordance with theoretical predictions. The internal velocity dispersion scales with the size of the cloud as $ $ for large clouds ($R pc $). For small clouds we find larger sigma compared to the power-law scaling, as seen in observations, which is due to supernova-driven turbulence. Almost all clouds are gravitationally unbound with the virial parameter scaling as $ vir $, which is slightly flatter compared to observed scaling but in agreement given the large scatter. We note that the cloud distribution towards the low-mass end is only complete if the more dilute gas is also refined, rather than only the collapsing regions.

Cloud detection from Himawari-8 spectral images using K-means++ clustering with the convolutional module

ABSTRACT Accurate and reliable cloud detection is one of the crucial preprocessing steps in satellite remote sensing. This paper adopts the spectral data in different channels observed by Himawari-8 geostationary satellite and proposes a cloud detection algorithm based on K-means++ clustering with a sharpening convolutional module (C-Kmeans++). The analysis found that the sum of the reflectance of channels 3 and 4, the brightness temperature of channel 15 and the difference of brightness temperature of channels 7 and 14 can be used as clustering features of the K-means++ algorithm. Before the clustering operation, the RS image is sharpened using the Laplacian operator to enhance the precision in the detection of thin clouds. A proper k value (k = 7) is obtained by the sum of squared error and silhouette coefficient. To evaluate the C-Kmeans++, the comparison of detection results using K-means++, the multi-spectral threshold, the k-nearest neighbour and the Vgg16+U-Net was carried out. The visible light image and Cloud-Aerosol Lidar with Orthogonal Polarization product were used as reference labels. It is found that RR, ER, HR and KSS of C-Kmeans++ are best throughout the day.

A Cloud Detection Algorithm Based on FY-4A/GIIRS Infrared Hyperspectral Observations

Cloud detection is an essential preprocessing step when using satellite-borne infrared hyperspectral sounders for data assimilation and atmospheric retrieval. In this study, we propose a cloud detection algorithm based solely on the sensitivity and detection characteristics of the FY-4A Geostationary Interferometric Infrared Sounder (GIIRS), rather than relying on other instruments. The algorithm consists of four steps: (1) combining observed radiation and clear radiance data simulated by the Community Radiative Transfer Model (CRTM) to identify clear fields of view (FOVs); (2) determining the number of clouds within adjacent 2 × 2 FOVs via a principal component analysis of observed radiation; (3) identifying whether there are large observed radiance differences between adjacent 2 × 2 FOVs to determine the mixture of clear skies and clouds; and (4) assigning adjacent 2 × 2 FOVs as a cloud cluster following the three steps above to select an appropriate classification threshold. The classification results within each cloud detection cluster were divided into the following categories: clear, partly cloudy, or overcast. The proposed cloud detection algorithm was tested using one month of GIIRS observations from May 2022 in this study. The cloud detection and classification results were compared with the FY-4A Advanced Geostationary Radiation Imager (AGRI)’s operational cloud mask products to evaluate their performance. The results showed that the algorithm’s performance is significantly influenced by the surface type. Among all-day observations, the highest recognition performance was achieved over the ocean, followed by land surfaces, with the lowest performance observed over deep inland water. The proposed algorithm demonstrated better clear sky recognition during the nighttime for ocean and land surfaces, while its performance was higher for partly cloudy and overcast conditions during the day. However, for inland water surfaces, the algorithm consistently exhibited a lower cloud recognition performance during both the day and night. Moreover, in contrast to the GIIRS’s Level 2 cloud mask (CLM) product, the proposed algorithm was able to identify partly cloudy conditions. The algorithm’s classification results departed slightly from those of the AGRI’s cloud mask product in areas with clear sky/cloud boundaries and minimal convective cloud coverage; this was attributed to the misclassification of clear sky as partly cloudy under a low-resolution situation. AGRI’s CLM products, temporally and spatially collocated to the GIIRS FOV, served as the reference value. The proportion of FOVs consistently classified as partly cloudy to the total number of partly cloudy FOVs was 40.6%. In comparison with the GIIRS’s L2 product, the proposed algorithm improved the identification performance by around 10%.

A hybrid cloud detection and cloud phase classification algorithm using classic threshold-based tests and extra randomized tree model

The detection of clouds from the latest-generation satellites is crucial for the accuracy of downstream remote sensing products. Classic threshold-based algorithms or machine-learning-based algorithms have limitations in all-day cloud detection, usually sacrificing measurements from visible bands to develop all-day cloud detections. It is difficult to find a day-night consistent cloud detection that takes advantage of all band measurements even if there is an abrupt absence of visible bands on one sides of the terminator. This study explains a cloud detection algorithm for the Cloud, Atmospheric Radiation, and renewal Energy application (CARE) datasets that synergistically uses threshold-based tests and an extra randomized tree (ERT) model to detect all-day clouds from Himawari-8 full-disk measurements. First, we developed a visible-to-infrared band daytime cloud detection algorithm (D algorithm) to generate high-precision training datasets, based on which an infrared band all-day cloud detection algorithm (DN algorithm) is built with an ERT model. Second, we merge the D and DN algorithm into CARE algorithm by weighting the results of the two algorithms differently according to day/night and land/sea conditions. In developing D algorithm, we use two-layer curved-surface thresholds from a look-up-table (LUT), that can effectively improve the dusk-dawn cloud detection accuracy over high latitudes and reduce the misclassification of clouds over sunglint regions. Validations against CALIPSO measurements show that the hit rate (HR) of CARE results (daytime cloudy HR 76.95%, clear HR 87.97%, nighttime cloudy HR 73.19%, clear HR 81.77%) outperform MODIS (daytime cloudy HR 77.35%, clear HR 66.09%, nighttime cloudy HR 78.08%, clear HR 59.72%) especially in nighttime clear sky detections, daytime CARE results show lower and higher HR in water/ice cloud detections than JAXA and MODIS results, nighttime CARE (water cloud HR 47.82%, ice cloud HR 93.59%) results show higher HR in both water and ice clouds than MODIS (water cloud HR 43.39%, ice cloud HR 74.41%). The comparison between the CARE and JAXA products shows high consistency during the daytime, and the CARE products have better cloud detection in the sunglint center and in dusk-dawn scenarios.

АНАЛІЗ ПІДХОДІВ ДО ВИДАЛЕННЯ ХМАРНОСТІ НА КОСМІЧНИХ ЗНІМКАХ ДИСТАНЦІЙНОГО ЗОНДУВАННЯ ЗЕМЛІ

The modern development of space technologies and remote sensing creates unique opportunities for solving problems in many areas, including the military. Remote sensing imagery often plays a key role in decision-making at all levels of military command, so one of the most important tasks in this context is cloud detection and extraction. This is an important stage of remote sensing data processing aimed at reconstructing information hidden by clouds. The article is devoted to the analysis of different approaches to cloud removal and improvement of the data quality. The approaches based on the use of various image processing algorithms (traditional approaches) have certain limitations associated with the frequent loss of useful information. Special attention is paid to deep learning methods, which have gained popularity in solving cloud removal problems. Deep Neural Networks show great potential for recovering information on satellite images that is hidden by clouds. This paper discusses various Deep Neural Networks architectures, such as convolutional neural networks, conditional generative adversarial networks, and their modifications. Their advantages and disadvantages are also considered. The use of such methods is more accurate and efficient compared to traditional image processing methods, as neural networks can adapt to various conditions and types of images. The analyzed disadvantages of fusing purely optical data led to the conclusion that the best approach to solving the problem of removing clouds from satellite images would be to combine optical and radar data. Despite the complexity of such an approach, it can show the greatest efficiency in solving the problem considered in this article. The challenges and prospects for further improvement of cloud removal methods on satellite images are considered. In particular, the use of artificial intelligence algorithms for automatic cloud detection and removal, as well as the need to create standardized methods for comparing and evaluating the effectiveness of different approaches. Keywords: satellite imagery; remote sensing; cloud cover; neural networks.

Improving Cloud Detection in WFV Images Onboard Chinese GF-1/6 Satellite

We have developed an algorithm for cloud detection in Chinese GF-1/6 satellite multispectral images, allowing us to generate cloud masks at the pixel level. Due to the lack of shortwave infrared and thermal infrared bands in the Chinese GF-1/6 satellite, bright land surfaces and snow are frequently misclassified as clouds. To mitigate this issue, we utilized MODIS standard snow data products for reference data to determine the presence of snow cover in the images. Subsequently, our algorithm was utilized to correct misclassifications in snow-covered mountainous regions. The experimental area selected was the perpetually snow-covered Western mountains in the United States. The results indicate the accurate labeling of extensive snow-covered areas, achieving an overall cloud detection accuracy of over 91%. Our algorithm enables users to easily determine whether pixels are affected by cloud contamination, effectively improving accuracy in annotating data quality and greatly facilitating subsequent data retrieval and utilization.

Polar Cloud Detection of FengYun-3D Medium Resolution Spectral Imager II Imagery Based on the Radiative Transfer Model

The extensive existence of high-brightness ice and snow underlying surfaces in polar regions presents notable complexities for cloud detection in remote sensing imagery. To elevate the accuracy of cloud detection in polar regions, a novel polar cloud detection algorithm is proposed in this paper. Employing the MOD09 surface reflectance product, we compiled a database of monthly composite surface reflectance in the shortwave infrared bands specific to polar regions. Through the forward simulation of the correlation between the apparent reflectance and surface reflectance across diverse conditions using the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) radiative transfer model, we established a dynamic cloud detection model for the shortwave infrared channels. In contrast to a machine learning algorithm and the widely used MOD35 cloud product, the algorithm introduced in this study demonstrates enhanced congruence with the authentic cloud distribution within cloud products. It precisely distinguishes between the cloudy and clear-sky pixels, achieving rates surpassing 90% for both, while maintaining an error rate and a missing rate each under 10%. The algorithm yields positive results for cloud detection in polar regions, effectively distinguishing between ice, snow, and clouds. It provides robust support for comprehensive and long-term cloud detection efforts in polar regions.

MoTI: A Multi-Stage Algorithm for Moving Object Identification in SLAM.

Simultaneous localization and mapping (SLAM) algorithms are widely applied in fields such as autonomous driving and target tracking. However, the effect of moving objects on localization and mapping remains a challenge in natural dynamic scenarios. To overcome this challenge, this paper proposes an algorithm for dynamic point cloud detection that fuses laser and visual identification data, the multi-stage moving object identification algorithm (MoTI). The MoTI algorithm consists of two stages: rough processing and precise processing. In the rough processing stage, a statistical method is employed to preliminarily detect dynamic points based on the range image error of the point cloud. In the precise processing stage, the radius search strategy is used to statistically test the nearest neighbor points. Next, visual identification information and point cloud registration results are fused using a method of statistics and information weighting to construct a probability model for identifying whether a point cloud cluster originates from a moving object. The algorithm is integrated into the front-end of the LOAM system, which significantly improves the localization accuracy. The MoTI algorithm is evaluated on an actual indoor dynamic environment and several KITTI datasets, and the results demonstrate its ability to accurately detect dynamic targets in the background and improve the localization accuracy of the robot.

Research on Progress of Forest Fire Monitoring with Satellite Remote Sensing

With satellite remote sensing technology blooming, satellite remote sensing has become a common tool to detect forest fires, and played an important role in forest fire monitoring. This paper sort the research status and progress on satellite remote sensing monitoring for forest fires to provide directions and insights for subsequent research and applications. Through reviewing the literature on satellite remote sensing monitoring for forest fires, we present satellites and sensors for forest fire monitoring, describe forest fire monitoring methods through brightness temperature detection and smoke detection, and summarize current problems of satellite remote sensing monitoring of forest fires. Despite forest fire satellite remote sensing monitoring algorithms are becoming increasingly mature, it is not without problems such as slow migration of cloud detection algorithms, difficulties in unifying spatial and temporal characteristics, and difficulties in detecting small fires and low-temperature fires. Finally, in response to the problems identified, we list some recommendations with a view to providing useful references for future research on forest fire monitoring with satellite remote sensing.

Cloud detection algorithm based on point by point refinement

In order to limit the interference of cloud noise on ground scene information, cloud detection has been a hot issue in research on remote sensing image processing. Cloud detection labels the clouds in remote sensing images at the pixel level. The majority of early cloud detection systems rely on manually created feature and threshold segmentation with limited generalizability. Remote sensing cloud detection based on deep learning has improved in accuracy and speed thanks to the quick development of convolutional neural networks, but it is still unable to satisfy practical application requirements when dealing with sceneries with variable cloud block size and sparse distribution. To this end, this study proposes a cloud detection algorithm based on point-by-point refinement based on the idea of coarse to fine. Specifically, firstly, the residual module is introduced in the U-Net network to extract more features; secondly, the point-by-point refinement module is designed to filter out the areas in the remote sensing images where the clouds are easily detected wrongly for optimization and re-prediction, and then produce finer-grained and more accurate cloud detection results. The quantitative and qualitative experiments validate the effectiveness of the proposed method.

Identification of Paddy Croplands and Its Stages Using Remote Sensors: A Systematic Review.

Rice is a staple food that feeds nearly half of the world's population. With the population of our planet expected to keep growing, it is crucial to carry out accurate mapping, monitoring, and assessments since these could significantly impact food security, climate change, spatial planning, and land management. Using the PRISMA systematic review protocol, this article identified and selected 122 scientific articles (journals papers and conference proceedings) addressing different remote sensing-based methodologies to map paddy croplands, published between 2010 and October 2022. This analysis includes full coverage of the mapping of rice paddies and their various stages of crop maturity. This review paper classifies the methods based on the data source: (a) multispectral (62%), (b) multisource (20%), and (c) radar (18%). Furthermore, it analyses the impact of machine learning on those methodologies and the most common algorithms used. We found that MODIS (28%), Sentinel-2 (18%), Sentinel-1 (15%), and Landsat-8 (11%) were the most used sensors. The impact of Sentinel-1 on multisource solutions is also increasing due to the potential of backscatter information to determine textures in different stages and decrease cloud cover constraints. The preferred solutions include phenology algorithms via the use of vegetation indices, setting thresholds, or applying machine learning algorithms to classify images. In terms of machine learning algorithms, random forest is the most used (17 times), followed by support vector machine (12 times) and isodata (7 times). With the continuous development of technology and computing, it is expected that solutions such as multisource solutions will emerge more frequently and cover larger areas in different locations and at a higher resolution. In addition, the continuous improvement of cloud detection algorithms will positively impact multispectral solutions.

GF-1/6 Satellite Pixel-by-Pixel Quality Tagging Algorithm

The Landsat and Sentinel series satellites contain their own quality tagging data products, marking the source image pixel by pixel with several specific semantic categories. These data products generally contain categories such as cloud, cloud shadow, land, water body, and snow. Due to the lack of mid-wave and thermal infrared bands, the accuracy of traditional cloud detection algorithm is unstable when facing Chinese Gaofen-1/6 (GF-1/6) data. Moreover, it is challenging to distinguish clouds from snow. In order to produce GF-1/6 satellite pixel-by-pixel quality tagging data products, this paper builds a training sample set of more than 100,000 image pairs, primarily using Sentinel-2 satellite data. Then, we adopt the Swin Transformer model with a self-attention mechanism for GF-1/6 satellite image quality tagging. Experiments show that the model’s overall accuracy reaches the level of Fmask v4.6 with more than 10,000 training samples, and the model can distinguish between cloud and snow correctly. Our GF-1/6 quality tagging algorithm can meet the requirements of the “Analysis Ready Data (ARD) Technology Research for Domestic Satellite” project.

A Cloud Detection Neural Network Approach for the Next Generation Microwave Sounder Aboard EPS MetOp-SG A1

This work presents an algorithm based on a neural network (NN) for cloud detection to detect clouds and their thermodynamic phase using spectral observations from spaceborne microwave radiometers. A standalone cloud detection algorithm over the ocean and land has been developed to distinguish clear sky versus ice and liquid clouds from microwave sounder (MWS) observations. The MWS instrument—scheduled to be onboard the first satellite of the Eumetsat Polar System Second-Generation (EPS-SG) series, MetOp-SG A1—has a direct inheritance from advanced microwave sounding unit A (AMSU-A) and the microwave humidity sounder (MHS) microwave instruments. Real observations from the MWS sensor are not currently available as its launch is foreseen in 2024. Thus, a simulated dataset of atmospheric states and associated MWS synthetic observations have been produced through radiative transfer calculations with ERA5 real atmospheric profiles and surface conditions. The developed algorithm has been validated using spectral observations from the AMSU-A and MHS sounders. While ERA5 atmospheric profiles serve as references for the model development and its validation, observations from AVHRR cloud mask products provide references for the AMSU-A/MHS model evaluation. The results clearly show the NN algorithm’s high skills to detect clear, ice and liquid cloud conditions against a benchmark. In terms of overall accuracy, the NN model features 92% (88%) on the ocean and 87% (85%) on land, for the MWS (AMSU-A/MHS)-simulated dataset, respectively.

Cloud Top Thermodynamic Phase from Synergistic Lidar-Radar Cloud Products from Polar Orbiting Satellites: Implications for Observations from Geostationary Satellites

The cloud thermodynamic phase is a crucial parameter to understand the Earth’s radiation budget, the hydrological cycle, and atmospheric thermodynamic processes. Spaceborne active remote sensing such as the synergistic radar-lidar DARDAR product is considered the most reliable method to determine cloud phase; however, it lacks large-scale observations and high repetition rates. These can be provided by passive instruments such as SEVIRI aboard the geostationary Meteosat Second Generation (MSG) satellite, but passive remote sensing of the thermodynamic phase is challenging and confined to cloud top. Thus, it is necessary to understand to what extent passive sensors with the characteristics of SEVIRI are expected to provide a relevant contribution to cloud phase investigation. To reach this goal, we collect five years of DARDAR data to model the cloud top phase (CTP) for MSG/SEVIRI and create a SEVIRI-like CTP through an elaborate aggregation procedure. Thereby, we distinguish between ice (IC), mixed-phase (MP), supercooled (SC), and warm liquid (LQ). Overall, 65% of the resulting SEVIRI pixels are cloudy, consisting of 49% IC, 14% MP, 13% SC, and 24% LQ cloud tops. The spatial resolution has a significant effect on the occurrence of CTP, especially for MP cloud tops, which occur significantly more often at the lower SEVIRI resolution than at the higher DARDAR resolution (9%). We find that SC occurs most frequently at high southern latitudes, while MP is found mainly in both high southern and high northern latitudes. LQ dominates in the subsidence zones over the ocean, while IC occurrence dominates everywhere else. MP and SC show little seasonal variability apart from high latitudes, especially in the south. IC and LQ are affected by the shift of the Intertropical Convergence Zone. The peak of occurrence of SC is at −3 ∘C, followed by that for MP at −13 ∘C. Between 0 and −27 ∘C, the occurrence of SC and MP dominates IC, while below −27 ∘C, IC is the most frequent CTP. Finally, the occurrence of cloud top height (CTH) peaks lower over the ocean than over land, with MP, SC, and IC being undistinguishable in the tropics but with separated CTH peaks in the rest of the MSG disk. Finally, we test the ability of a state-of-the-art AI-based ice cloud detection algorithm for SEVIRI named CiPS (Cirrus Properties for SEVIRI) to detect cloud ice. We confirm previous evaluations with an ice detection probability of 77.1% and find a false alarm rate of 11.6%, of which 68% are due to misclassified cloud phases. CiPS is not sensitive to ice crystals in MP clouds and therefore not suitable for the detection of MP clouds but only for fully glaciated (i.e., IC) clouds. Our study demonstrates the need for the development of dedicated cloud phase distinction algorithms for all cloud phases (IC, LQ, MP, SC) from geostationary satellites.

Convolutional Neural Network-Driven Improvements in Global Cloud Detection for Landsat 8 and Transfer Learning on Sentinel-2 Imagery

A stable and reliable cloud detection algorithm is an important step of optical satellite data preprocessing. Existing threshold methods are mostly based on classifying spectral features of isolated individual pixels and do not contain or incorporate the spatial information. This often leads to misclassifications of bright surfaces, such as human-made structures or snow/ice. Multi-temporal methods can alleviate this problem, but cloud-free images of the scene are difficult to obtain. To deal with this issue, we extended four deep-learning Convolutional Neural Network (CNN) models to improve the global cloud detection accuracy for Landsat imagery. The inputs are simplified as all discrete spectral channels from visible to short wave infrared wavelengths through radiometric calibration, and the United States Geological Survey (USGS) global Landsat 8 Biome cloud-cover assessment dataset is randomly divided for model training and validation independently. Experiments demonstrate that the cloud mask of the extended U-net model (i.e., UNmask) yields the best performance among all the models in estimating the cloud amounts (cloud amount difference, CAD = −0.35%) and capturing the cloud distributions (overall accuracy = 94.9%) for Landsat 8 imagery compared with the real validation masks; in particular, it runs fast and only takes about 41 ± 5.5 s for each scene. Our model can also actually detect broken and thin clouds over both dark and bright surfaces (e.g., urban and barren). Last, the UNmask model trained for Landsat 8 imagery is successfully applied in cloud detections for the Sentinel-2 imagery (overall accuracy = 90.1%) via transfer learning. These prove the great potential of our model in future applications such as remote sensing satellite data preprocessing.