4:30 PM - 4:45 PM
[ACG41-29] Evaluating Cloud Flag Product of Spaceborne Multispectral Imager Onboard GCOM-C Using CloudSat and CALIPSO over High-Latitude Region
Keywords:GCOM-C, Cloud, Imager, Machine learning
Cloud detection is one of the most critical elements in the Earth’s radiation budget and climate system. The role of clouds in the Earth’s energy balance is pivotal, as they reflect incoming solar radiation and absorb outgoing longwave radiation, which makes accurate cloud detection essential for understanding the Earth’s climate and improving climate models. However, clouds are still a major uncertainty in global warming predictions (IPCC AR6 2021). Despite advances in satellite-based cloud observation technologies, cloud detection remains a challenge, particularly in high-latitude regions, where snow and ice-covered surfaces create ambiguity in cloud identification.
The Global Change Observation Mission-Climate (GCOM-C), launched by the Japan Aerospace Exploration Agency (JAXA), carries the Second-generation Global Imager (SGLI), a multispectral imager. One of the key products derived from the SGLI sensor is the Cloud Flag product (CLFG), which classifies each pixel as either clear sky or cloudy.
This study presents a comprehensive evaluation of the GCOM-C/SGLI Cloud Flag in the high-latitude regions of the Northern Hemisphere. To assess the accuracy of cloud detection, we compare CLFG data with cloud detection information from a spaceborne cloud radar (CloudSat/CPR) and atmospheric lidar (CALIPSO/CALIOP), which are active sensors enable to provide more accurate cloud data because they can capture vertical profiles. The study focuses particularly on the latest version of the CLFG product (Ver. 3), which incorporates a deep neural network (DNN) model alongside the original algorithm (called CLAUDIA).
A large dataset of matchup pairs was created, containing over 600,000 pairs of GCOM-C and CloudSat-CALIPSO data points. This dataset covers the period from October 2018 to June 2019 and was matched in both space and time, with a maximum time difference of 10 minutes. The comparison between CLFG Ver. 2 and Ver. 3 are conducted in various statistical values such as cloud fraction, several statistical metrics (Overall Accuracy, Hit Rate, Success Ratio, Bias Score, Threat Score) with divided datasets, that is, during the day or at night and over land or ocean. Moreover, spatial distributions of statistical metrics are analyzed.
Overall results show that CLFG Ver. 2 and Ver. 3 showed an overall improvement in cloud detection accuracy of 2%, from 70.4% in Ver. 2 to 72.4% in Ver. 3. This mainly attributes the significant improvement over Land during the day by 8.1%. This is because the DNN method identifies clouds which is missed in CLFG Ver.2. Over oceans during the day, GCOM-C Ver. 3 has an overall accuracy of 78.4%, the highest for all categories, though the improvement over Ver. 2 is negligible.
On the other hand, the study also highlights conditions where challenges remain. The cloudy bias of GCOM-C CLFG is observed at night both over land and ocean, and this issue remains unresolved in Ver. 3. Cloud detection using only thermal infrared channels available at night is challenging even with the DNN method. Also, cloud detection over Greenland is difficult due to the distinctive optical properties of the ice sheet in Greenland.
This study also investigated a key feature of the CLFG product, which is uncertainty information provided through the Clear Confidence Level (CCL). The CCL value ranges from 0 to 1, where values closer to 0 indicate high confidence in cloud detection, and values closer to 1 indicate high confidence in clear sky identification. By adjusting the threshold of the CCL, users can select data based on the desired level of accuracy. The results show that Ver. 3 enhances the feature, allowing for more flexible and accurate use of the cloud flag product.
In conclusion, this study demonstrates the improvement in cloud detection provided by Ver. 3 of the GCOM-C/SGLI Cloud Flag product, focusing on high-latitude regions where cloud detection is particularly challenging. The implementation of the DNN model in Ver. 3 has led to significant improvements over land during daytime, and enhancement of the uncertainty information which allows users to flexibly extract data to suit their needs. However, challenges remain at night and ice- and snow-covered surface in Greenland.
The Global Change Observation Mission-Climate (GCOM-C), launched by the Japan Aerospace Exploration Agency (JAXA), carries the Second-generation Global Imager (SGLI), a multispectral imager. One of the key products derived from the SGLI sensor is the Cloud Flag product (CLFG), which classifies each pixel as either clear sky or cloudy.
This study presents a comprehensive evaluation of the GCOM-C/SGLI Cloud Flag in the high-latitude regions of the Northern Hemisphere. To assess the accuracy of cloud detection, we compare CLFG data with cloud detection information from a spaceborne cloud radar (CloudSat/CPR) and atmospheric lidar (CALIPSO/CALIOP), which are active sensors enable to provide more accurate cloud data because they can capture vertical profiles. The study focuses particularly on the latest version of the CLFG product (Ver. 3), which incorporates a deep neural network (DNN) model alongside the original algorithm (called CLAUDIA).
A large dataset of matchup pairs was created, containing over 600,000 pairs of GCOM-C and CloudSat-CALIPSO data points. This dataset covers the period from October 2018 to June 2019 and was matched in both space and time, with a maximum time difference of 10 minutes. The comparison between CLFG Ver. 2 and Ver. 3 are conducted in various statistical values such as cloud fraction, several statistical metrics (Overall Accuracy, Hit Rate, Success Ratio, Bias Score, Threat Score) with divided datasets, that is, during the day or at night and over land or ocean. Moreover, spatial distributions of statistical metrics are analyzed.
Overall results show that CLFG Ver. 2 and Ver. 3 showed an overall improvement in cloud detection accuracy of 2%, from 70.4% in Ver. 2 to 72.4% in Ver. 3. This mainly attributes the significant improvement over Land during the day by 8.1%. This is because the DNN method identifies clouds which is missed in CLFG Ver.2. Over oceans during the day, GCOM-C Ver. 3 has an overall accuracy of 78.4%, the highest for all categories, though the improvement over Ver. 2 is negligible.
On the other hand, the study also highlights conditions where challenges remain. The cloudy bias of GCOM-C CLFG is observed at night both over land and ocean, and this issue remains unresolved in Ver. 3. Cloud detection using only thermal infrared channels available at night is challenging even with the DNN method. Also, cloud detection over Greenland is difficult due to the distinctive optical properties of the ice sheet in Greenland.
This study also investigated a key feature of the CLFG product, which is uncertainty information provided through the Clear Confidence Level (CCL). The CCL value ranges from 0 to 1, where values closer to 0 indicate high confidence in cloud detection, and values closer to 1 indicate high confidence in clear sky identification. By adjusting the threshold of the CCL, users can select data based on the desired level of accuracy. The results show that Ver. 3 enhances the feature, allowing for more flexible and accurate use of the cloud flag product.
In conclusion, this study demonstrates the improvement in cloud detection provided by Ver. 3 of the GCOM-C/SGLI Cloud Flag product, focusing on high-latitude regions where cloud detection is particularly challenging. The implementation of the DNN model in Ver. 3 has led to significant improvements over land during daytime, and enhancement of the uncertainty information which allows users to flexibly extract data to suit their needs. However, challenges remain at night and ice- and snow-covered surface in Greenland.