FACT: Interactive Deep Learning Platform and Multi-source Data Integration for Improved Soil Moisture Forecasting

The overarching goal of this proposal is to develop a deep learning computational platform that integrates multi-scale, multi-sensor soil moisture observations, a seasonal climate model, and user input in real-time, and provides useful climate information for agricultural planning on sub-seasonal to seasonal timescales. The proposal builds on the several years of collaborative research conducted by PD Kumar and Co-PDs Lee and Rangwala in the areas of hydroclimate predictability and cyberinfrastructure (Govindaraju et al. 2009; Kumar and Merwade 2009; Kumar et al. 2010; Kumar et al. 2014d), and drought processes and monitoring (Dewes et al. 2017; Rangwala et al. 2016). Our recent work includes (1) an in-depth investigation of a new climate process the soil moisture reemergence for improving skill of the soil moisture forecast (Kumar et al. 2019a); (2) a survey gauging the need for software-defined, storage-based data infrastructure and climate model simulation containers (Lee and Kumar 2016); and (3) development of high-resolution datasets to assess soil moisture state (Rangwala et al. 2019). The central hypothesis of the proposed research is that the Big Data revolution and latest advances in Earth System Modeling can help minimize the agricultural loss due to drought by providing skillful forecast information at sub-seasonal to seasonal time scales. The average cost per drought in the U.S. from 1980 to 2013 was estimated to be $9.5 billion dollars (Smith and Matthews 2015). The 2012 drought was the most extensive U.S. drought on record from 1895 to present and resulted in $30 billion economic losses, largely from the agricultural sector (Hoerling et al. 2014; Rippey 2015). While droughts are naturally occurring events that can be exacerbated by climate change (Ault et al. 2018; Cook et al. 2016), a marriage of the Big Data revolution with Earth System Modeling can deliver resources that can help the communities better prepare and plan for an upcoming drought. Deep Learning: The deep learning (DL) method has potential to provide scientific breakthroughs in soil moisture forecasting (Shen et al. 2018). The DL method transforms a set of input layers into several hidden layers with each successive hidden layer representing a higher level of data abstraction using general purpose learning to reach the final output layer (LeCun et al. 2015). The DL model can learn location invariant transferrable knowledge from the soil moisture-monitoring network observation data and apply the learned knowledge to forecast soil moisture at user location (Bruzzone and Marconcini 2010; Chu et al. 2013; Gong et al. 2013). We hypothesize that the location invariant properties can be derived from the high-resolution biophysical data (e.g., topography, land use and soil types, and low-resolution seasonal climate forecast). By combining the latest advances in earth system modeling and deep learning, we propose to develop the Next Generation Interactive Soil Moisture Forecasting System (NG-ISMFS) that is agile, smart, and adaptable to meet the needs of stakeholders and deliver the latest science at user fingertips. We have detailed the underlying new algorithms and technologies for the proposed NG-ISMFS in the Approach section. We have three short-term objectives that will contribute toward building the NG-ISMFS: Objective 1: Develop new algorithms and methods for integrating soil moisture data from different sources, including user input, climate model forecast data, and remotely sensed soil moisture observations to produce the best value product for the U.S. agriculture industry (Lead: PD Kumar) Objective 2: Build a scalable big data infrastructure and deep learning analytics platform for real-time and interactive soil moisture forecast applications (Lead co-PD: Lee) Objective 3: Develop new or improved forecast attributes at the interface of Human-Technology-Data interactions through workshops with stakeholders. (Lead: co-PD: Rangwala)