A Deep Learning Approach for Cloud Masking in GeoNEX Data Products

Kyle Kabasares
NASA Ames Research Center

Stevenson 1300
12:00 PM

Accurately identifying clouds in satellite images is important for studying the Earth’s surface. Traditional methods, such as radiative transfer models and basic machine learning techniques, often struggle to process complex satellite data from the GOES-R system. In this work, I developed a deep learning-based approach to improve cloud masking by combining spatial and temporal information. I built a U-Net convolutional neural network (CNN) to analyze images pixel by pixel, classifying each pixel as CLEAR, PARTLY CLOUDY, or CLOUDY. The model was trained using a set of carefully selected satellite images and achieved nearly 88% accuracy on test and validation data. However, some challenges remain, such as distinguishing clouds from shadows and snow-covered mountains, which can lead to misclassification. Future improvements will focus on optimizing input data, refining the labeling process, and designing better methods to measure model reliability.