580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can machine learning and deep learning improve cloud type classification and characterization in satellite and model data?
This theme focuses on the application of advanced machine learning (ML) and deep learning (DL) methods, including convolutional neural networks (CNNs), to classify and characterize cloud types from satellite observations and model outputs. It addresses the challenge of identifying traditional cloud genera and mesoscale cloud morphologies using automated techniques on low-resolution and multi-spectral data, which is crucial for understanding cloud feedbacks in climate models, improving weather forecasting, and enhancing Earth observation capabilities.
Key finding: Developed a supervised deep convolutional ANN capable of determining WMO cloud genera from low-resolution top-of-atmosphere radiation data (2.5° spatial resolution) with application to satellite data (CERES) and climate model... Read more
Cloud Mesoscale Cellular Classification and Diurnal Cycle Using a Convolutional Neural Network (CNN)
Key finding: Developed a CNN for pixel-level classification of marine stratocumulus mesoscale cellular convection (open and closed cells) using visible and infrared geo-stationary satellite SEVIRI imagery with 30 min temporal resolution.... Read more
Key finding: Established a framework to train and compare state-of-the-art deep neural networks (two-stage R-CNN and U-net variants) for cloud detection and cloud coverage prediction based on omnidirectional ground-based sky camera images... Read more
Key finding: Proposed ACLNet, a hybrid model combining EfficientNet-B0 backbone with à trous spatial pyramid pooling (ASPP), a global attention module, and integration of K-means clustering for precise cloud boundary segmentation in... Read more
Key finding: Designed an on-board cloud coverage detection system for small satellites employing a lightweight convolutional neural network optimized for deployment on commercial off-the-shelf microcontroller platforms. The system... Read more
2. What methodologies and datasets support accurate cloud detection, segmentation, and classification from satellite and ground-based remote sensing data?
This theme addresses the development, evaluation, and comparison of cloud detection methods, including classical approaches, machine learning, and hybrid algorithms applied to multispectral satellite data and ground-based imaging systems. It highlights the significance of large, quality-controlled datasets for training and benchmarking, discusses challenges such as thin cloud and cloud shadow discrimination, and covers algorithmic strategies to handle varying spatial-temporal resolutions and sensor characteristics, which are essential for improving operational cloud masks and related climate and weather applications.
Key finding: Provided a systematic multi-faceted review of cloud detection and tracking methods based on satellite and ground-based images, classifying techniques into statistical, physical retrieval-based, machine learning, and hybrid... Read more
Key finding: Surveyed threshold-based algorithms, classical image parameterization, and recent machine learning (notably CNN) methods for detecting cloud presence, thin/thick cloud discrimination, and snow/cloud confusion in satellite... Read more
Key finding: Developed CloudSEN12, an unprecedented large-scale, globally distributed, multi-temporal Sentinel-2 dataset containing 49,400 image patches with multiple annotation types distinguishing clear sky, thick/thin clouds, and cloud... Read more
Key finding: Introduced a Cloud Detection Algorithm-Generating (CDAG) method using a hyperspectral AVIRIS dataset to simulate multispectral sensors and devise reliable cloud detection algorithms based on spectral differences between... Read more
Key finding: Proposed a scalable, landmark-specific cloud detection framework for geostationary satellite imagery using an ensemble of dedicated support vector machines trained under season and illumination stratifications. Applied to MSG... Read more
3. How do advanced algorithms and analytical models improve cloud property retrieval and classification from multispectral and lidar remote sensing data to enhance climate and environmental applications?
This research area explores the development of physically informed analytical models, machine learning-based classifications, and retrieval algorithms for cloud optical, physical, and vertical profile properties using multispectral imagers and lidar data. It encompasses challenges of vertical structure inference, low-density airborne laser scanning point cloud classification, cloud type retrieval for climate constraint, and aerosol-cloud discrimination, supporting improved climate model evaluation, ecological monitoring, and operational Earth system observation.
Key finding: Developed an optimal estimation-based retrieval algorithm operating on ESA's EarthCARE MSI multi-spectral visible and thermal infrared bands to jointly infer cloud optical thickness, effective radius, and cloud top height... Read more
Key finding: Employed CSU RAMS cloud model simulations and empirical orthogonal function (EOF) analysis to identify four dominant triangle-shaped vertical patterns of liquid water content and effective radius in stratocumulus profiles,... Read more
Machine Learning-Based Supervised Classification of Point Clouds Using Multiscale Geometric Features
Key finding: Evaluated eight supervised machine learning classifiers for semantic labeling of 3D point clouds from airborne laser scanning across multiple spatial scales using multiscale geometric features based on eigenvalue-derived... Read more
Key finding: Presented a novel classification workflow integrating geometric features and K-means clustering of LiDAR intensity (GIK procedure) for distinguishing low-density airborne laser scanner point clouds of shrubs from similarly... Read more
Key finding: Showcased machine learning methodologies utilizing Cloud-Aerosol Transport System (CATS) space-based lidar data onboard the ISS to detect and classify vertically resolved cloud and aerosol layers. Addressed challenges of low... Read more