Machine Learning (ML) enabled forecasts derived from Carbon, Silicate, and Nitrogen Ecosystem (CoSiNE) model generated climatology

Christopher Wood1, Richard W Gould Jr2, Bradley Penta3, Sergio DeRada3, Sean McCarthy2 and Gregory P Behm4, (1)United States, (2)US Naval Research Laboratory, Stennis Space Center, MS, United States, (3)Naval Research Laboratory, Stennis Space Center, MS, United States, (4)High Performance Computing Modernization Program, Productivity Enhancement, Technology Transfer, and Training (PETTT), Vicksburg, MS, United States
Abstract:
The NASA-funded Adaptive Ecosystem Climatology (AEC) project at NRL blends static satellite and model fields with near-real time satellite imagery to create merged 3-dimensional (3D) products. Using Machine Learning (ML) techniques, we first generated multivariate relationships between the ecosystem model (CoSiNE) and Moderate Resolution Imaging Spectroradiometer (MODIS-Aqua) ocean color fields, using a training data set based on 30-year model and 20-year image climatologies. Using a combination of depth slicing and fusing of satellite surface data to existing climatological data we developed a predictive solution for fifteen physical, chemical, and biological model output variables. The goal is to use near-real time surface satellite imagery in conjunction with the predictive relationships, to adjust the static fields to provide better estimates of the 3D model fields. Using the input of geospatial meta-data, chlorophyll (CHL), and sea surface temperature (SST) derived from daily satellite imagery, the predictive capability successfully generated outputs commensurate with the CoSiNE model climatology and Subject Matter Expert (SME) expectations. AEC’s ML based predictive capability is targeting integration with NOAA web-based data servers. AEC’s Artificial Intelligence (AI) could be utilized to support a spectrum of efforts ranging from localized ecosystem support functions to fisheries model inputs and harmful algal bloom support. We will present methods of data collection and encapsulation, recognizing concerns related to data quantity / density, training methodologies, and consideration for ensembles and tools used.