GOESCloudPredictions is an excursion in applied Machine Learning. It looked to solve one core issue: Improving cloud forecasts and interpolating image information between snapshots of the GOES-16 satellite (which occur every 15 minutes)

Files of Interest: Final Project Report: Technical report on the entirety of the project, its workings, and its findings. Geostationary Cloud Tracking Spring 2024 - Presentation of the model's findings and decision making rationale Old: Interim model's report, findings, and other documentation Data: MP4 - Example GIF format of predictions PNGs - Processed, masked GOES-16 images to show greyscaled cloud cover. Used directly as input for CNN, ConvLSTM models Cloud Predictions: Convolution with Separation: Kernel matrix is decomposed into two smaller, one-dimensional vectors through two steps. Convolution without Separation: Convlution is performed with full kernel matrix where kernel is simply slided over input images Convolutional: Base CNN model application and checkpoints Frame Interpolation: Contains framework for CNN-based image interpolation model and checkpoints. LSTM Model: Contains the ConvLSTM model which maintained the highest training/testing accuracy. Cloud Masking: Background subtraction preprocessing code used to visualize GOES images, apply visual bands to simulate a cloud cover, contrast clouds from background using background subtraction across different times of day, and convert them into PNG format (Data) for CNN/ConvLSTM model input

Contact [email protected] for any further inquiries, thank you!