Single-Station Forecasting from Deep Learning Methods

Prediction of surface visibility and cloud ceilings remains a significant challenge for not just military weather services, but the operational meteorology community as a whole. Another specific challenge to the military is developing robust forecasting applications that operate in communications-limited environments in deployed locations. The research seeks to develop a deep learning trained model that will provide a skillful forecast at point locations across the globe with flexible input, ranging from the full suite of numerical weather prediction products, down to a 12 hour observing record at an austere location. To accomplish this goal, this research will utilize the global, quality controlled surface observation dataset from the 14^th Weather Squadron, the Global Ensemble Forecast System Reforecast version 2 and high resolution global topography and land cover data. This research will experiment with a combination of unsupervised and supervised machine learning applications to test the effectiveness of clustering stations by various methods, including climatology and/or local topography and land cover prior to utilizing modern long/short term memory (LSTM) and convolutional neural network deep learning methods. The goal is to develop a proof of concept for a lightweight, global deployable application for use by operational military forecasters to develop skillful, single station forecasts of operationally significant parameters.