CyberSEES:Type 2: Precipitation Estimation from Multi-Source Information using Advanced Machine Learning

University Of California-Irvine, Irvine CA

Investigators

Abstract

This project will develop a cyber-enabled, data-driven modeling system that will use the vast amount of earth and environmental observational data to estimate precipitation, with the goal to further freshwater resource sustainability. One of the major objectives of this research is to innovatively apply advanced machine learning techniques to predict complex natural phenomena. Identification of current machine learning algorithms' representational and computational limitations when applied to earth and environmental data that are required for accurately estimating precipitation will also be explored. Specifically, this project will explore image-based feature extraction techniques and so-called "deep belief" models for interpreting important features within weather and climate data to estimate and predict precipitation. Features to be investigated include appearance, texture, shape, dynamics, and regional weather and climate signatures. The project will use data from ground-based radar, satellite-based sensors, and Numerical Weather Prediction (NWP) models, as well as physical characteristics from land surface datasets. Image-based extraction techniques will produce feature maps, which in turn will be used as input into a Deep Boltzmann Machine (DBM) model, a modern version of neural networks. A DBM represents a probability model over a collection of visible units and hidden units and produces as output a target variable (in this case, precipitation). This model is particularly suitable for the proposed modeling approach and for large-scale data. Once this system is developed, we will utilize a wide variety of earth science data to provide accurate, high-resolution global precipitation estimates. Verification methods to be used for evaluating precipitation estimates will include general statistics, precipitation intensity distribution, and regional analysis methods used by the atmospheric and hydrological sciences. The project will focus on verification over the United States, where high-resolution ground-based radar data are available. In recent decades, extreme flooding and droughts have become more frequent and severe. Changing patterns in precipitation, which are attributed to climate variability, are responsible for these hydrologic extremes and contribute to uncertainties in freshwater resource management and planning. In the face of the planet's growing population and stresses on water resources, it is important to minimize these uncertainties and the social impacts of these natural hazards. These goals can only be accomplished through accurate precipitation measurements and forecasts. Satellite platforms and advanced numerical models produce massive amounts of global high temporal and spatial resolution data that can be used for this purpose, but analyzing these data remains a challenge. Recent innovations in computational sciences and machine learning have extended our capability to harvest from remotely-sensed data critical information that is essential to understanding cloud-precipitation systems. This project will adapt and improve satellite-based precipitation estimation algorithms by using computational science, statistical modeling techniques (machine learning), remote-sensing observations, and numerical models to develop cyber-enabled modeling systems that can effectively analyze the massive amount of observational data to improve the global estimation of precipitation at high spatial and temporal resolutions.