Deep-learning Based Cloud Fraction Retrieval and Rainy Cloud Classification Using Satellite Remote Sensing Data




Liu, Qian

Journal Title

Journal ISSN

Volume Title



Cloud is the carrier of precipitation and an important atmospheric factor that influence both the long-term climate change and short-term weather dynamics of the earth (Molinari and Dudek, 1992; Rossow and Schiffer, 1999; Hamilton, 2006). It substantially affects Earth’s energy budget by reflecting solar radiation back to space and by restricting emission of thermal radiation to space (Ramanathan et al. 1989), and is important regulator of the climate and earth-atmosphere system (Sassen et al., 2008). The profiles and types of clouds are also crucial parameters of global climate models (GCM) and numerical weather predictions (NWP). They are considered to be the largest uncertainty in the analysis and prediction of climate change, owning to the difference between climate models and observational datasets over the area where clouds occur (Dufresne and Bony, 2008). Cloud features such as cloud fractions and precipitation capabilities have significant different of impact on the earth system and are related to various natural disasters. The lack of complete knowledge concerning the complex interactions among clouds, circulation, and climate hinders our ability to simulate the Earth’s climate correctly (Daloz et al., 2018). Furthermore, in the condition of climate change, many cloud-related procedures and changes occur in scales that is smaller than the climate model grids. It is impossible to include these sub-scale processes and their response to increasing temperature in current climate change simulations. As a conclusion, a comprehensive investigation of cloud distribution and classification in high resolution is essential for the research and analysis of the entire Earth system. Regarding the importance of cloud fraction in the performance of climate models and lack of cloud fraction estimation of hyperspectral IR sounders, a deep neural network model is created to retrieve the cloud fraction within the field of view (FOV) of Cross-track Infrared Sounder (CrIS). To reduce the model input factors without losing spectral information thus increase the retrieval efficiency, principal component transformation is performed on the original CrIS spectrum and sensitivity tests are conducted to determine the best performing combination of PCs as model predictors. During the training procedure, the best-performed iteration and epoch numbers are also tested to avoid over fitting. In general, the cloud fraction retrieved from the proposed DNN model are consistent with truth values calculated from the VIIRS cloud mask product, resulting in a low Mean Square Error (MSE) of 0.021 and a high Pearson correlation coefficient (R) of 0.924. Regarding the challenge of low spatiotemporal resolutions of microwave (MV) precipitation products and the relative low accuracy of infrared (IR) products, this study combines the information and takes advantages from both MW and IR data. A deep-learning-based rainy cloud detection and classification framework is developed using ABI spectrum as input predictors and IMERG precipitation estimates as learning target. With the high spatial-temporal resolution of ABI images, the proposed system will be of high performance in real-time regional and local precipitation monitoring. And to include full coverage of precipitation characteristics of the study area, IMERG is used as truth instead of discretely distributed ground observations. The assessment parameters indicate that the proposed models produce relatively accurate results with a critical success index (CSI) of 0.71 and a probability of detection (POD) of 0.86 for rainy cloud detection, and CSI of 0.58 and a POD of 0.72 for convective clouds delineation. Regarding the relatively low efficiencies of traditional algorithms in processing large remote sensing data, the study utilizes deep learning method to classify rainy cloud types and estimate cloud fractions. With the flourish of Artificial intelligence (AI) and big data techniques, AI methods such as machine and deep learning have been broadly adopted to investigate geospatial and climatological phenomena, as well as predict natural disasters, which triggers a new concept, GeoAI. Although traditional machine learning methods have shown their capability and potential in precipitation detection and monitoring, deep learning (DL) approaches are more accurate in processing big data with large volume and various features (LeCun et al., 2015), such as remote sensing images which contain an abundance of spatial, temporal and spectral information. The study proposes deep-learning based cloud classification and fraction retrieval framework using GEO and LEO satellite data, satellite precipitation product and cloud mask product to detect and classify the rainy clouds of Advanced Baseline Imager (ABI) into stratiform and convective, and estimate the cloud fractions in the field of views (FOVs) of hyperspectral sounder. This research is innovative for the following reasons: 1) cloud fractions of hyper-spectral sounders have rarely been addressed but are important for most climate and weather forecast models, this study retrieve the real cloud fraction in each field of view (FOV) of hyper-spectral sounder; 2) most passive microwave (PMW) based satellite precipitation products are at low spatiotemporal resolutions and ground precipitation measurements are sparsely distributed, the study creates a high-resolution rainy cloud type product by combining PMW and geostationary satellite data, to support the precipitation disaster management; 3). After the models are successfully trained, the retrieval and detection results can be produced in high efficiency, which avoids the complex calculation in traditional algorithms; 4). The mature and advanced AI method in the computer science field is utilized to explore new application from the observations from the JPSS and GOES satellites.



Geography, Cloud fraction, Deep neural network, GOES-16 ABI, Infrared sounder, Principal component analysis, Rainy cloud